diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_0/index_tensor([7318, 7318])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_0/index_tensor([7318, 7318])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb9166c05b853366f754ba9a7fc3b286a5c61698 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_0/index_tensor([7318, 7318])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle close-up captures a scene of profound tranquility and minimalist elegance. A vibrant crimson silk scarf, its fabric so light and sheer that its fine weave is almost translucent, lies spread across the smooth, dark mahogany surface of a polished wooden table. Its high-gloss finish reflects light like a still, dark pool of water. A gentle, unseen draft of air infuses the scene with energy, delicately lifting the scarf from the table. The fabric captures the current, first fluttering at the edge, then gracefully swelling upward in a slow-motion ballet of soft, tumbling waves. The silk rises and falls in a fluid, organic movement, then slowly, almost reluctantly, falls back down to drape the wood in a new, graceful form. The entire scene is bathed in a soft, directional light, as if from a window outside the frame, casting gentle, elongated shadows on the scarf and highlighting the silk's sheen against the deep, rich grain of the table, creating a striking, beautiful contrast.", + "seed": 27899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_0/index_7318_seed_27899_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": -0.04629420423880219, + "vlm_logprob": -0.04629420423880219, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.04629420423880219, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09558174759149551 + ], + "opposite_logprob_tokens": [ + -2.4705817699432373 + ], + "logprob_reward": -0.09558174759149551, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09558174759149551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10095614194869995, + 0.03323259577155113, + 0.04882179945707321, + 0.02172059565782547, + 0.3103732466697693, + 0.17501284182071686, + 0.17180907726287842, + 0.03049209527671337, + 0.08101893961429596, + 0.026562660932540894 + ], + "credit_map": [ + 0.005833272356539965, + 0.01228029653429985, + 0.00892981979995966, + 0.015626342967152596, + 0.015044540166854858, + 0.009645512327551842, + 0.017323732376098633, + 0.01627263054251671, + 0.002490910002961755, + 0.0037363083101809025, + 0.0030846898443996906, + 0.004962288774549961, + 0.003963279537856579, + 0.004715020302683115, + 0.005710545461624861, + 0.004569556564092636, + 0.003947726916521788, + 0.007693275343626738, + 0.0032440898939967155, + 0.005595394875854254, + 0.006773368455469608, + 0.007529019843786955, + 0.008705375716090202, + 0.005333544686436653, + 0.0011270950781181455, + 0.004474626388400793, + 0.0018640839261934161, + 0.0029322863556444645, + 0.003386198543012142, + 0.0018877409165725112, + 0.0033647119998931885, + 0.0026838521007448435, + 0.0069014267064630985, + 0.09869149327278137, + 0.005591280292719603, + 0.03015502169728279, + 0.046691883355379105, + 0.00866179820150137, + 0.08823522180318832, + 0.025445111095905304, + 0.0065362839959561825, + 0.018775109201669693, + 0.01216677762567997, + 0.028593363240361214, + 0.018791833892464638, + 0.01792941614985466, + 0.048631373792886734, + 0.02358868159353733, + 0.005257295444607735, + 0.01521285343915224, + 0.01147602777928114, + 0.0282560084015131, + 0.025257892906665802, + 0.01334608905017376, + 0.05459247902035713, + 0.01841042935848236, + 0.0014681066386401653, + 0.0029940633103251457, + 0.002426175167784095, + 0.00460549583658576, + 0.003157199826091528, + 0.003199682105332613, + 0.00720031326636672, + 0.0054410588927567005, + 0.003980047069489956, + 0.004420903977006674, + 0.005808386951684952, + 0.029547255486249924, + 0.005876267794519663, + 0.004193501081317663, + 0.017492834478616714, + 0.00969974510371685, + 0.0015204971423372626, + 0.001721506123431027, + 0.0020739021711051464, + 0.009005779400467873, + 0.0023397214245051146, + 0.0019356196280568838, + 0.004901304375380278, + 0.0030643304344266653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.09558174759149551 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296396702528 + ], + "opposite_logprob_tokens": [ + -2.237963914871216 + ], + "logprob_reward": -0.11296396702528, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296396702528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08127692341804504, + 0.049300599843263626, + 0.08793371915817261, + 0.025159602984786034, + 0.38280239701271057, + 0.10925453156232834, + 0.12576280534267426, + 0.10598660260438919, + 0.015243593603372574, + 0.01727921888232231 + ], + "credit_map": [ + 0.004088074434548616, + 0.0070531112141907215, + 0.008517167530953884, + 0.020638776943087578, + 0.008016843348741531, + 0.005207208916544914, + 0.016615424305200577, + 0.01114032231271267, + 0.0023905413690954447, + 0.004766117315739393, + 0.003651972860097885, + 0.009493611752986908, + 0.00501970062032342, + 0.005901524797081947, + 0.010799131356179714, + 0.007277997210621834, + 0.006226030178368092, + 0.0110591109842062, + 0.005226261913776398, + 0.008281225338578224, + 0.009523173794150352, + 0.012720378115773201, + 0.020337658002972603, + 0.01455987524241209, + 0.00179840880446136, + 0.0034778877161443233, + 0.002799698617309332, + 0.0035622119903564453, + 0.00280071166343987, + 0.0043939645402133465, + 0.0033908255863934755, + 0.002935894299298525, + 0.00837924424558878, + 0.12366581708192825, + 0.007012241519987583, + 0.037060532718896866, + 0.06032383069396019, + 0.009409778751432896, + 0.10583054274320602, + 0.031120389699935913, + 0.004371039103716612, + 0.01092155184596777, + 0.008439301513135433, + 0.017020829021930695, + 0.012302981689572334, + 0.012684207409620285, + 0.029216056689620018, + 0.014298560097813606, + 0.003922892268747091, + 0.011508525349199772, + 0.008740008808672428, + 0.021605439484119415, + 0.018978137522935867, + 0.007301324047148228, + 0.040133703500032425, + 0.013572776690125465, + 0.003374092746526003, + 0.01009279303252697, + 0.006322452332824469, + 0.02296125516295433, + 0.009214751422405243, + 0.0063943201676011086, + 0.035678599029779434, + 0.011948337778449059, + 0.000927783374208957, + 0.0009717017528600991, + 0.0011939163086935878, + 0.005023435223847628, + 0.0012994298012927175, + 0.0009135109721682966, + 0.003008452942594886, + 0.0019053631694987416, + 0.0008780166390351951, + 0.0010952985612675548, + 0.0013337347190827131, + 0.005556347314268351, + 0.0017705248901620507, + 0.0015860575949773192, + 0.0031211012974381447, + 0.0019381383899599314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.11296396702528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046386257745325565 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.0046386257745325565, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046386257745325565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14537666738033295, + 0.039423007518053055, + 0.06591910123825073, + 0.03168068453669548, + 0.3279927372932434, + 0.19424298405647278, + 0.12093713879585266, + 0.022700881585478783, + 0.02596229314804077, + 0.025764448568224907 + ], + "credit_map": [ + 0.0075263893231749535, + 0.012818069197237492, + 0.01593063585460186, + 0.034020502120256424, + 0.012493353337049484, + 0.011828567832708359, + 0.02798057161271572, + 0.0227785874158144, + 0.0026356843300163746, + 0.005313803441822529, + 0.0028356246184557676, + 0.005066396668553352, + 0.00491879740729928, + 0.0060643418692052364, + 0.006631631404161453, + 0.005956726148724556, + 0.006265179254114628, + 0.00780529947951436, + 0.005949642509222031, + 0.009060356765985489, + 0.006844463292509317, + 0.009146985597908497, + 0.010903229005634785, + 0.0099439462646842, + 0.0015445390017703176, + 0.004990427289158106, + 0.0034216982312500477, + 0.004881701897829771, + 0.0037381176371127367, + 0.0032766261138021946, + 0.005825404543429613, + 0.0040021720342338085, + 0.008321872912347317, + 0.1127152293920517, + 0.00474284403026104, + 0.027911076322197914, + 0.05630142614245415, + 0.008374053984880447, + 0.08514218032360077, + 0.02448405511677265, + 0.006966860964894295, + 0.018536942079663277, + 0.017143169417977333, + 0.034009065479040146, + 0.0165705569088459, + 0.020170841366052628, + 0.052123259752988815, + 0.02872229553759098, + 0.003659452311694622, + 0.01012461818754673, + 0.009853485971689224, + 0.020385006442666054, + 0.016206607222557068, + 0.009840156883001328, + 0.03740343824028969, + 0.013464367017149925, + 0.0009560322505421937, + 0.002435081172734499, + 0.0016672109486535192, + 0.0035927186254411936, + 0.00225981418043375, + 0.002751413267105818, + 0.005515629891306162, + 0.003522980958223343, + 0.0014131894567981362, + 0.002111803274601698, + 0.002967842621728778, + 0.005853078793734312, + 0.0028681685216724873, + 0.002307558199390769, + 0.0037083232309669256, + 0.004732328932732344, + 0.0013734659878537059, + 0.002135690301656723, + 0.0022862767800688744, + 0.00784766674041748, + 0.0027330894954502583, + 0.001893661916255951, + 0.004445591010153294, + 0.0030490050558000803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0046386257745325565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191891998052597 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014191891998052597, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191891998052597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055594056844711304, + 0.02132471278309822, + 0.11809585243463516, + 0.04132262244820595, + 0.3080933690071106, + 0.2856731414794922, + 0.06339229643344879, + 0.03794115409255028, + 0.033771466463804245, + 0.034791313111782074 + ], + "credit_map": [ + 0.004305013921111822, + 0.007790104486048222, + 0.0044489288702607155, + 0.006481654942035675, + 0.009372665546834469, + 0.0057675507850945, + 0.008885075338184834, + 0.008543061092495918, + 0.0015324538107961416, + 0.0031738532707095146, + 0.002341283019632101, + 0.0028666756115853786, + 0.002489575417712331, + 0.003239138750359416, + 0.0025680121034383774, + 0.003113720566034317, + 0.009096282534301281, + 0.015718087553977966, + 0.008098638616502285, + 0.011827072128653526, + 0.013940482400357723, + 0.018999772146344185, + 0.02417682856321335, + 0.01623869501054287, + 0.0024731261655688286, + 0.0060081202536821365, + 0.00332028535194695, + 0.004626373760402203, + 0.005610367748886347, + 0.008883586153388023, + 0.005574410781264305, + 0.004826352931559086, + 0.008873607963323593, + 0.10907428711652756, + 0.005968289449810982, + 0.023997444659471512, + 0.05583846941590309, + 0.010148114524781704, + 0.07379110157489777, + 0.020402034744620323, + 0.01154831051826477, + 0.027052124962210655, + 0.022383056581020355, + 0.04429818317294121, + 0.02894013188779354, + 0.03541421517729759, + 0.07756786793470383, + 0.03846925497055054, + 0.0019157935166731477, + 0.005679313093423843, + 0.005069767124950886, + 0.011261522769927979, + 0.009228385053575039, + 0.0030106797348707914, + 0.02167486399412155, + 0.005551969166845083, + 0.0014170915819704533, + 0.004365512169897556, + 0.002879364648833871, + 0.008317270316183567, + 0.002885869238525629, + 0.0027723959647119045, + 0.01096934825181961, + 0.004334300756454468, + 0.0018285434925928712, + 0.0024592585396021605, + 0.0025950714480131865, + 0.011425591073930264, + 0.002914619166404009, + 0.0019560835789889097, + 0.005679928231984377, + 0.004912371281534433, + 0.0015392301138490438, + 0.0025366374757140875, + 0.002486695069819689, + 0.011009406298398972, + 0.004028992727398872, + 0.0030859617982059717, + 0.006304432637989521, + 0.0037999588530510664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.014191891998052597 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094788804650307 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094788804650307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094788804650307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10517061501741409, + 0.05161786451935768, + 0.058740418404340744, + 0.03157560154795647, + 0.4137445092201233, + 0.08397518843412399, + 0.09082634747028351, + 0.03882819414138794, + 0.07399381697177887, + 0.05152743309736252 + ], + "credit_map": [ + 0.006702141370624304, + 0.010407493449747562, + 0.012330872938036919, + 0.026677845045924187, + 0.010303664021193981, + 0.008959158323705196, + 0.01641717366874218, + 0.013372262939810753, + 0.003117094049230218, + 0.006303353700786829, + 0.004690766334533691, + 0.006954549811780453, + 0.0056099365465343, + 0.004726037383079529, + 0.009881232865154743, + 0.010334892198443413, + 0.002727656392380595, + 0.006039382424205542, + 0.002616686513647437, + 0.005384537857025862, + 0.005307798273861408, + 0.00601536687463522, + 0.017395781353116035, + 0.013253206387162209, + 0.001553946640342474, + 0.007540970109403133, + 0.0019911667332053185, + 0.004142299294471741, + 0.005076276604086161, + 0.004148103296756744, + 0.004390112590044737, + 0.0027327232528477907, + 0.010432311333715916, + 0.12406585365533829, + 0.016628243029117584, + 0.043709155172109604, + 0.05793562904000282, + 0.010443862527608871, + 0.11455179005861282, + 0.035977695137262344, + 0.0037309566978365183, + 0.009086442179977894, + 0.007495439145714045, + 0.011022076942026615, + 0.011765559203922749, + 0.012965764850378036, + 0.016490446403622627, + 0.011418500915169716, + 0.0027569872327148914, + 0.009045871905982494, + 0.006942515727132559, + 0.015051414258778095, + 0.013761992566287518, + 0.007486939430236816, + 0.02710244432091713, + 0.008678183890879154, + 0.0010478571057319641, + 0.003330652602016926, + 0.0021399392280727625, + 0.008951977826654911, + 0.0034068503882735968, + 0.002105885650962591, + 0.012761411257088184, + 0.005083617754280567, + 0.004098086152225733, + 0.0038540721870958805, + 0.00437376881018281, + 0.028870288282632828, + 0.0048990859650075436, + 0.00320435524918139, + 0.016422288492321968, + 0.008271872997283936, + 0.0027507655322551727, + 0.003620221046730876, + 0.0035574412904679775, + 0.01881292462348938, + 0.004491571336984634, + 0.0029489584267139435, + 0.00940657313913107, + 0.005938977934420109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.004094788804650307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a scene of profound tranquility and minimalist elegance. A vibrant crimson silk scarf, its fabric so light and sheer that its fine weave is almost translucent, lies spread across the smooth, dark mahogany surface of a polished wooden table. Its high-gloss finish reflects light like a still, dark pool of water. A gentle, unseen draft of air infuses the scene with energy, delicately lifting the scarf from the table. The fabric captures the current, first fluttering at the edge, then gracefully swelling upward in a slow-motion ballet of soft, tumbling waves. The silk rises and falls in a fluid, organic movement, then slowly, almost reluctantly, falls back down to drape the wood in a new, graceful form. The entire scene is bathed in a soft, directional light, as if from a window outside the frame, casting gentle, elongated shadows on the scarf and highlighting the silk's sheen against the deep, rich grain of the table, creating a striking, beautiful contrast.", + "seed": 27900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_0/index_7318_seed_27900_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": -0.025315883010625838, + "vlm_logprob": -0.025315883010625838, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.025315883010625838, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011379575356841087 + ], + "opposite_logprob_tokens": [ + -5.386379718780518 + ], + "logprob_reward": -0.011379575356841087, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.011379575356841087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10449894517660141, + 0.04972202330827713, + 0.07149676978588104, + 0.0962291732430458, + 0.39335140585899353, + 0.033930703997612, + 0.042530499398708344, + 0.022959472611546516, + 0.02394857071340084, + 0.16133250296115875 + ], + "credit_map": [ + 0.014387859031558037, + 0.016064446419477463, + 0.013212178833782673, + 0.01666450686752796, + 0.008314131759107113, + 0.007420104928314686, + 0.019454339519143105, + 0.008981375955045223, + 0.006207514554262161, + 0.008396865800023079, + 0.005272343289107084, + 0.009635874070227146, + 0.005666705779731274, + 0.003070975886657834, + 0.007323635742068291, + 0.0041481079533696175, + 0.007915475405752659, + 0.011610985733568668, + 0.007262760773301125, + 0.009700549766421318, + 0.00820252113044262, + 0.006704416591674089, + 0.013402102515101433, + 0.006697960663586855, + 0.007331245578825474, + 0.007072826381772757, + 0.009229392744600773, + 0.03506572172045708, + 0.004234387539327145, + 0.01630338467657566, + 0.010550164617598057, + 0.006442046258598566, + 0.039487048983573914, + 0.05719446390867233, + 0.05107414722442627, + 0.07221240550279617, + 0.04409737139940262, + 0.019720477983355522, + 0.08015580475330353, + 0.02940966747701168, + 0.0049464707262814045, + 0.008356671780347824, + 0.0031771743670105934, + 0.0030909141059964895, + 0.003484359011054039, + 0.0031713496427983046, + 0.0033438573591411114, + 0.004359906539320946, + 0.005830895621329546, + 0.004673399496823549, + 0.006120454054325819, + 0.0059664687141776085, + 0.00525775458663702, + 0.002877553692087531, + 0.008157988078892231, + 0.003645989578217268, + 0.003088579513132572, + 0.0027325949631631374, + 0.0028139285277575254, + 0.003006732789799571, + 0.00233278120867908, + 0.002046537585556507, + 0.003448604140430689, + 0.0034897138830274343, + 0.0032433923333883286, + 0.004793263040482998, + 0.002391873160377145, + 0.0028183162212371826, + 0.002978638745844364, + 0.0015341192483901978, + 0.003308818442746997, + 0.0028801492881029844, + 0.012349233031272888, + 0.03143022581934929, + 0.015334597788751125, + 0.03142257034778595, + 0.014876797795295715, + 0.006033875979483128, + 0.03160005435347557, + 0.018285145983099937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.011379575356841087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900442183017731 + ], + "opposite_logprob_tokens": [ + -2.5790045261383057 + ], + "logprob_reward": -0.07900442183017731, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900442183017731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09088148176670074, + 0.05541130527853966, + 0.112190842628479, + 0.11955155432224274, + 0.34410926699638367, + 0.021191956475377083, + 0.054741568863391876, + 0.016320019960403442, + 0.02577361650764942, + 0.15982839465141296 + ], + "credit_map": [ + 0.014307878911495209, + 0.014897217974066734, + 0.011453058570623398, + 0.012676541693508625, + 0.009173616766929626, + 0.008451953530311584, + 0.011771288700401783, + 0.008149920962750912, + 0.005991274490952492, + 0.010422874242067337, + 0.006475935224443674, + 0.008443029597401619, + 0.005365204066038132, + 0.004512069281190634, + 0.007584681268781424, + 0.0066162412986159325, + 0.013581854291260242, + 0.01535871997475624, + 0.01132262870669365, + 0.017971130087971687, + 0.010934283025562763, + 0.007461280096322298, + 0.023056115955114365, + 0.012504825368523598, + 0.029924539849162102, + 0.006938031408935785, + 0.007258065510541201, + 0.02229050174355507, + 0.007137159816920757, + 0.026781544089317322, + 0.007857036776840687, + 0.011364668607711792, + 0.03342623636126518, + 0.04973506182432175, + 0.04183601960539818, + 0.06122094392776489, + 0.04010298103094101, + 0.01581963710486889, + 0.0727582722902298, + 0.029210126027464867, + 0.0030995861161500216, + 0.0025073818396776915, + 0.002997534116730094, + 0.0024879034608602524, + 0.00222376873716712, + 0.001985100330784917, + 0.004110269248485565, + 0.0017804148374125361, + 0.006869950331747532, + 0.008990800008177757, + 0.006703960243612528, + 0.006290721241384745, + 0.006855058018118143, + 0.004160023760050535, + 0.006607597693800926, + 0.008263456635177135, + 0.002211017534136772, + 0.0026917376089841127, + 0.0022638170048594475, + 0.002676567994058132, + 0.0016021557385101914, + 0.0014543944271281362, + 0.0022253962233662605, + 0.0011949329636991024, + 0.0037768578622490168, + 0.005728008691221476, + 0.0028383710887283087, + 0.0034756166860461235, + 0.0020528545137494802, + 0.002789711346849799, + 0.0032254059333354235, + 0.0018867886392399669, + 0.012271338142454624, + 0.031187325716018677, + 0.015248280018568039, + 0.029948962852358818, + 0.01540787611156702, + 0.005992891266942024, + 0.03193642944097519, + 0.017835302278399467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.07900442183017731 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04653961956501007, + 0.06637337803840637, + 0.4087444841861725, + 0.0628712847828865, + 0.03229304775595665, + 0.013319249264895916, + 0.016943437978625298, + 0.032050974667072296, + 0.048827361315488815, + 0.272037148475647 + ], + "credit_map": [ + 0.008757420815527439, + 0.006654968019574881, + 0.006121926475316286, + 0.007222327869385481, + 0.004361424595117569, + 0.003581242635846138, + 0.005522022023797035, + 0.004318287596106529, + 0.00883878581225872, + 0.014399340376257896, + 0.006322912871837616, + 0.008244913071393967, + 0.006312804762274027, + 0.004331317730247974, + 0.008128493092954159, + 0.0097948107868433, + 0.03578516095876694, + 0.045364487916231155, + 0.046105269342660904, + 0.08803217858076096, + 0.03463044390082359, + 0.028082365170121193, + 0.09087888151407242, + 0.039865702390670776, + 0.014186359010636806, + 0.004109064117074013, + 0.004221268929541111, + 0.007390961050987244, + 0.004502126481384039, + 0.01730666495859623, + 0.004148502834141254, + 0.007006337400525808, + 0.003537736600264907, + 0.004947178531438112, + 0.004378949292004108, + 0.005594407208263874, + 0.0034710837062448263, + 0.0015668136766180396, + 0.006470640655606985, + 0.0023262372706085443, + 0.0020100276451557875, + 0.0013396760914474726, + 0.0016609570011496544, + 0.0016867243684828281, + 0.0015781604452058673, + 0.0013382536126300693, + 0.002094184048473835, + 0.0016112657031044364, + 0.002479074290022254, + 0.0030673944856971502, + 0.0027826763689517975, + 0.0024203599896281958, + 0.0015573136042803526, + 0.0010973465396091342, + 0.0021066819317638874, + 0.001432590652257204, + 0.004325421527028084, + 0.005062622018158436, + 0.004357953555881977, + 0.004202499054372311, + 0.003004617290571332, + 0.0034817582927644253, + 0.00458529544994235, + 0.0030308051500469446, + 0.007990200072526932, + 0.008784563280642033, + 0.004220643546432257, + 0.005256542470306158, + 0.004832230042666197, + 0.0050345598720014095, + 0.0062104398384690285, + 0.006498183589428663, + 0.02046963945031166, + 0.050412941724061966, + 0.025997880846261978, + 0.0534079447388649, + 0.02481439895927906, + 0.013118664734065533, + 0.052628662437200546, + 0.031187022104859352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.016062384471297264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.09564980119466782, + 0.062408607453107834, + 0.1266413927078247, + 0.08795619010925293, + 0.18791598081588745, + 0.031151199713349342, + 0.09771675616502762, + 0.04987059906125069, + 0.10088232159614563, + 0.1598072350025177 + ], + "credit_map": [ + 0.021298861131072044, + 0.019839633256196976, + 0.010452799499034882, + 0.008725631050765514, + 0.008398921228945255, + 0.006684926338493824, + 0.00983382761478424, + 0.010415189899504185, + 0.011064802296459675, + 0.014961488544940948, + 0.00653453404083848, + 0.008332273922860622, + 0.004573488142341375, + 0.0039290678687393665, + 0.006853500381112099, + 0.006159449461847544, + 0.011624127626419067, + 0.016862325370311737, + 0.014867992140352726, + 0.027140779420733452, + 0.010298045352101326, + 0.00924622267484665, + 0.025050649419426918, + 0.011551239527761936, + 0.019483176991343498, + 0.006558975670486689, + 0.006557763088494539, + 0.00803576409816742, + 0.00624490063637495, + 0.024486947804689407, + 0.0061028082855045795, + 0.010485849343240261, + 0.01826155185699463, + 0.028497586026787758, + 0.02084968611598015, + 0.02973928675055504, + 0.02089092880487442, + 0.011106438934803009, + 0.040328510105609894, + 0.018241990357637405, + 0.0035190279595553875, + 0.003700618864968419, + 0.0055550080724060535, + 0.004180074203759432, + 0.0026747803203761578, + 0.0022037022281438112, + 0.0050932010635733604, + 0.004224785603582859, + 0.012507762759923935, + 0.017774119973182678, + 0.010805832222104073, + 0.010604876093566418, + 0.012375143356621265, + 0.008274921216070652, + 0.010470441542565823, + 0.014903652481734753, + 0.005182934459298849, + 0.005792147014290094, + 0.005466364324092865, + 0.00833386741578579, + 0.005369645543396473, + 0.005489019677042961, + 0.00880993902683258, + 0.005426680203527212, + 0.013321121223270893, + 0.013514209538698196, + 0.011156016029417515, + 0.01590236835181713, + 0.01059088483452797, + 0.01030124444514513, + 0.014957278966903687, + 0.011139197275042534, + 0.012752627953886986, + 0.02848951704800129, + 0.015826553106307983, + 0.029028987511992455, + 0.016787214204669, + 0.0065491944551467896, + 0.03157712519168854, + 0.018795998767018318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "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\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944667384028435 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944667384028435, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944667384028435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2489296793937683, + 0.04218930006027222, + 0.30129146575927734, + 0.03148694708943367, + 0.0868639126420021, + 0.03179202601313591, + 0.03219272568821907, + 0.031307514756917953, + 0.04142390936613083, + 0.1525225192308426 + ], + "credit_map": [ + 0.06044597923755646, + 0.06816636025905609, + 0.021222524344921112, + 0.02500765211880207, + 0.01891157031059265, + 0.01273779384791851, + 0.021570948883891106, + 0.02086682990193367, + 0.006243162322789431, + 0.0066600097343325615, + 0.00487565016373992, + 0.007502423133701086, + 0.0038183829747140408, + 0.0027049777563661337, + 0.00548709649592638, + 0.004897595848888159, + 0.026036614552140236, + 0.034231774508953094, + 0.03409126400947571, + 0.06290652602910995, + 0.02669224701821804, + 0.020699573680758476, + 0.06830266863107681, + 0.028330780565738678, + 0.00841253437101841, + 0.0023516248911619186, + 0.003005732549354434, + 0.004297226667404175, + 0.0018651856807991862, + 0.0055149020627141, + 0.0025512396823614836, + 0.0034885001368820667, + 0.00970600824803114, + 0.01363022718578577, + 0.012272887863218784, + 0.01507897861301899, + 0.009075210429728031, + 0.004001378081738949, + 0.017279427498579025, + 0.005819784477353096, + 0.0036318469792604446, + 0.005914418958127499, + 0.004393334500491619, + 0.003755206009373069, + 0.0035197476390749216, + 0.002711863722652197, + 0.0034534987062215805, + 0.004412107635289431, + 0.003963824827224016, + 0.009412200190126896, + 0.003324309829622507, + 0.0037238709628582, + 0.0030809652525931597, + 0.0015816790983080864, + 0.003909291699528694, + 0.0031965805683285, + 0.004221293143928051, + 0.0040246471762657166, + 0.0048040603287518024, + 0.003856667084619403, + 0.003068156773224473, + 0.002597565995529294, + 0.004601045977324247, + 0.004134076181799173, + 0.006079846061766148, + 0.0057976143434643745, + 0.003999315667897463, + 0.005892651621252298, + 0.004404515027999878, + 0.006341238971799612, + 0.005767961964011192, + 0.0031407636124640703, + 0.0142229488119483, + 0.0310171227902174, + 0.013583768159151077, + 0.02598755992949009, + 0.015834204852581024, + 0.008382640779018402, + 0.027897443622350693, + 0.015596813522279263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.005944667384028435 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a scene of profound tranquility and minimalist elegance. A vibrant crimson silk scarf, its fabric so light and sheer that its fine weave is almost translucent, lies spread across the smooth, dark mahogany surface of a polished wooden table. Its high-gloss finish reflects light like a still, dark pool of water. A gentle, unseen draft of air infuses the scene with energy, delicately lifting the scarf from the table. The fabric captures the current, first fluttering at the edge, then gracefully swelling upward in a slow-motion ballet of soft, tumbling waves. The silk rises and falls in a fluid, organic movement, then slowly, almost reluctantly, falls back down to drape the wood in a new, graceful form. The entire scene is bathed in a soft, directional light, as if from a window outside the frame, casting gentle, elongated shadows on the scarf and highlighting the silk's sheen against the deep, rich grain of the table, creating a striking, beautiful contrast.", + "seed": 27901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_0/index_7318_seed_27901_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": -0.016781899519264697, + "vlm_logprob": -0.016781899519264697, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.016781899519264697, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01198507659137249 + ], + "opposite_logprob_tokens": [ + -5.2619853019714355 + ], + "logprob_reward": -0.01198507659137249, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01198507659137249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4252568185329437, + 0.027608884498476982, + 0.024455798789858818, + 0.19316110014915466, + 0.09278624504804611, + 0.07261057943105698, + 0.10332335531711578, + 0.018314303830266, + 0.017705664038658142, + 0.024777183309197426 + ], + "credit_map": [ + 0.08872317522764206, + 0.05556686967611313, + 0.03781794384121895, + 0.12087950855493546, + 0.02923201210796833, + 0.02379867248237133, + 0.03462876006960869, + 0.03460989147424698, + 0.002593257464468479, + 0.0034611301962286234, + 0.0032858254853636026, + 0.002223447198048234, + 0.005461341235786676, + 0.003484213026240468, + 0.003502732142806053, + 0.003596941474825144, + 0.0019915406592190266, + 0.0055376808159053326, + 0.002174741355702281, + 0.0020473708864301443, + 0.002934335032477975, + 0.001358897890895605, + 0.005756895057857037, + 0.002654339885339141, + 0.01740487664937973, + 0.052644822746515274, + 0.014976102858781815, + 0.017272181808948517, + 0.020137939602136612, + 0.011994092725217342, + 0.04721047356724739, + 0.011520631611347198, + 0.009403732605278492, + 0.01177408080548048, + 0.009078188799321651, + 0.011324734427034855, + 0.018362075090408325, + 0.010669424198567867, + 0.01253895740956068, + 0.009635050781071186, + 0.00649902131408453, + 0.01735910028219223, + 0.009118301793932915, + 0.005531012546271086, + 0.008460275828838348, + 0.003942166455090046, + 0.017358968034386635, + 0.0043417406268417835, + 0.010442334227263927, + 0.012960807420313358, + 0.02012171410024166, + 0.010492200031876564, + 0.016457339748740196, + 0.013408737257122993, + 0.009119146503508091, + 0.010321089066565037, + 0.002027560491114855, + 0.0031283129937946796, + 0.002481842180714011, + 0.002255604835227132, + 0.002508376259356737, + 0.00144966843072325, + 0.0022756371181458235, + 0.0021873046644032, + 0.0013559692306444049, + 0.002314957557246089, + 0.0023223680909723043, + 0.0016892586136236787, + 0.0035823273938149214, + 0.0018630058038979769, + 0.002074707765132189, + 0.0025030719116330147, + 0.002600154373794794, + 0.004728388506919146, + 0.004545008763670921, + 0.0023852756712585688, + 0.0033307860139757395, + 0.0019626275170594454, + 0.002646609442308545, + 0.0025783360470086336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.01198507659137249 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048677146434783936 + ], + "opposite_logprob_tokens": [ + -3.0486772060394287 + ], + "logprob_reward": -0.048677146434783936, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048677146434783936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2607848644256592, + 0.06750759482383728, + 0.23568658530712128, + 0.14799098670482635, + 0.034692876040935516, + 0.05462726950645447, + 0.10060180723667145, + 0.02849249169230461, + 0.03591139614582062, + 0.033704083412885666 + ], + "credit_map": [ + 0.0533406138420105, + 0.033385761082172394, + 0.03125225380063057, + 0.0704515278339386, + 0.019120197743177414, + 0.015071138739585876, + 0.018727656453847885, + 0.019435718655586243, + 0.005731360521167517, + 0.009639553725719452, + 0.009684056043624878, + 0.006352235563099384, + 0.012688599526882172, + 0.005303818732500076, + 0.006697754841297865, + 0.011410211212933064, + 0.01674107275903225, + 0.08715209364891052, + 0.009265202097594738, + 0.013405618257820606, + 0.018051529303193092, + 0.009338884614408016, + 0.06755281984806061, + 0.014179364778101444, + 0.010996037162840366, + 0.04176205396652222, + 0.009941041469573975, + 0.009378181770443916, + 0.0135332141071558, + 0.010187400504946709, + 0.04302820563316345, + 0.009164857678115368, + 0.003587417770177126, + 0.003933385014533997, + 0.005111352074891329, + 0.005467921029776335, + 0.004344678483903408, + 0.0031418651342391968, + 0.00551292859017849, + 0.0035933279432356358, + 0.006283264607191086, + 0.008260906673967838, + 0.011556441895663738, + 0.0068389130756258965, + 0.00794608797878027, + 0.0037578477058559656, + 0.007044714409857988, + 0.0029390971176326275, + 0.009346835315227509, + 0.011766387149691582, + 0.01961328648030758, + 0.010449807159602642, + 0.018974309787154198, + 0.014184749685227871, + 0.009945361874997616, + 0.006321064196527004, + 0.002448291750624776, + 0.005953643005341291, + 0.003969154786318541, + 0.003937224857509136, + 0.003500351682305336, + 0.0025372831150889397, + 0.002848932519555092, + 0.00329761253669858, + 0.004280571825802326, + 0.006839867681264877, + 0.005398275330662727, + 0.003136401530355215, + 0.006192138884216547, + 0.0025025634095072746, + 0.003596122609451413, + 0.003965457435697317, + 0.0038578358944505453, + 0.0077475327998399734, + 0.0040522548370063305, + 0.004116242751479149, + 0.005546402186155319, + 0.0029646605253219604, + 0.0028960835188627243, + 0.0025230715982615948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.048677146434783936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780935011804104 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780935011804104, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780935011804104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1283770650625229, + 0.13365410268306732, + 0.13213400542736053, + 0.24083082377910614, + 0.09991412609815598, + 0.05684041231870651, + 0.0554024837911129, + 0.05623016133904457, + 0.025408215820789337, + 0.07120862603187561 + ], + "credit_map": [ + 0.012907843105494976, + 0.020862724632024765, + 0.01319321058690548, + 0.018473392352461815, + 0.016087708994746208, + 0.016923977062106133, + 0.013782737776637077, + 0.016145464032888412, + 0.015095478855073452, + 0.018587378785014153, + 0.013086114078760147, + 0.00996660627424717, + 0.027925537899136543, + 0.016712334007024765, + 0.011183717288076878, + 0.021096937358379364, + 0.011015912517905235, + 0.03182922303676605, + 0.012016297318041325, + 0.01129723247140646, + 0.011989598162472248, + 0.012846197001636028, + 0.025355085730552673, + 0.015784453600645065, + 0.011252813041210175, + 0.0832470953464508, + 0.008676556870341301, + 0.013333075679838657, + 0.019772276282310486, + 0.010987565852701664, + 0.07924508303403854, + 0.01431635208427906, + 0.009982501156628132, + 0.008562403731048107, + 0.008514446206390858, + 0.01272619515657425, + 0.012244809418916702, + 0.016536034643650055, + 0.01610332541167736, + 0.01524441223591566, + 0.005633223336189985, + 0.008740796707570553, + 0.008957690559327602, + 0.006182205397635698, + 0.007672213949263096, + 0.004098373465240002, + 0.010903764516115189, + 0.004652145318686962, + 0.01093092281371355, + 0.006499856244772673, + 0.00851690024137497, + 0.0047730933874845505, + 0.007784429006278515, + 0.006650895345956087, + 0.006766369100660086, + 0.0034800185821950436, + 0.005876462906599045, + 0.007702245377004147, + 0.00494155939668417, + 0.003367580706253648, + 0.010540159419178963, + 0.009577721357345581, + 0.006242976523935795, + 0.007981456816196442, + 0.002776614623144269, + 0.005543103441596031, + 0.0035759657621383667, + 0.0024900208227336407, + 0.003909709397703409, + 0.0014893742045387626, + 0.0026192034129053354, + 0.003004224505275488, + 0.008298732340335846, + 0.008162562735378742, + 0.009132386185228825, + 0.005902973935008049, + 0.008627904579043388, + 0.008290630765259266, + 0.01001743134111166, + 0.01277600135654211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.009780935011804104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067337024956941605, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067337024956941605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3748277723789215, + 0.029052825644612312, + 0.059230607002973557, + 0.35618188977241516, + 0.016419844701886177, + 0.0588003471493721, + 0.023709848523139954, + 0.017106864601373672, + 0.05668879300355911, + 0.007981179282069206 + ], + "credit_map": [ + 0.08429012447595596, + 0.029430463910102844, + 0.029336869716644287, + 0.09923465549945831, + 0.028525860980153084, + 0.029917094856500626, + 0.038094379007816315, + 0.03599831834435463, + 0.0023763140197843313, + 0.0038983826525509357, + 0.0031094991136342287, + 0.0023095759097486734, + 0.006830514874309301, + 0.0033827389124780893, + 0.0032536741346120834, + 0.0038921243976801634, + 0.0036609736271202564, + 0.022886214777827263, + 0.002440392505377531, + 0.0028955950401723385, + 0.0051735276356339455, + 0.0030292130541056395, + 0.016360370442271233, + 0.0027843236457556486, + 0.0184632521122694, + 0.12430072575807571, + 0.015023134648799896, + 0.01887030154466629, + 0.029721826314926147, + 0.01146821491420269, + 0.12265807390213013, + 0.015676351264119148, + 0.001614039414562285, + 0.00178880977910012, + 0.001909147365950048, + 0.0019443450728431344, + 0.0030158651061356068, + 0.0016521405195817351, + 0.002454117638990283, + 0.0020413787569850683, + 0.006319799460470676, + 0.011440271511673927, + 0.00895613431930542, + 0.005666834302246571, + 0.007029365748167038, + 0.003712491365149617, + 0.012046318501234055, + 0.0036291321739554405, + 0.005753605626523495, + 0.002644769847393036, + 0.0029716298449784517, + 0.0023816123139113188, + 0.00241942354477942, + 0.002413972048088908, + 0.0029923240654170513, + 0.0021325107663869858, + 0.001703214249573648, + 0.0032886466942727566, + 0.002160902600735426, + 0.0018042089650407434, + 0.002934487769380212, + 0.0017912465846166015, + 0.001741936313919723, + 0.0016822206089273095, + 0.005637500435113907, + 0.00833133328706026, + 0.008392691612243652, + 0.005714741535484791, + 0.011035309173166752, + 0.0034153263550251722, + 0.006059475243091583, + 0.00810241512954235, + 0.0008650554227642715, + 0.0009048916981555521, + 0.0016398896696045995, + 0.0008016546489670873, + 0.0013644559076055884, + 0.000745009514503181, + 0.000777708541136235, + 0.000882513471879065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0067337024956941605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067326370626688 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067326370626688, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067326370626688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2007531374692917, + 0.04699341580271721, + 0.10709311813116074, + 0.10924983024597168, + 0.06726392358541489, + 0.14343775808811188, + 0.05221937969326973, + 0.20023523271083832, + 0.018460815772414207, + 0.05429335683584213 + ], + "credit_map": [ + 0.030946195125579834, + 0.025801420211791992, + 0.016433389857411385, + 0.046383488923311234, + 0.018368422985076904, + 0.014540270902216434, + 0.021005740389227867, + 0.02727421373128891, + 0.005511911120265722, + 0.006173396948724985, + 0.0057583642192184925, + 0.0037521778140217066, + 0.007715760264545679, + 0.004436441697180271, + 0.004493808373808861, + 0.009151558391749859, + 0.007025104481726885, + 0.036791495978832245, + 0.006251990329474211, + 0.005412884056568146, + 0.00799365434795618, + 0.005137583706527948, + 0.03043334372341633, + 0.008047067560255527, + 0.009360694326460361, + 0.015126446262001991, + 0.010220539756119251, + 0.011356205679476261, + 0.01486626360565424, + 0.022154610604047775, + 0.014891876839101315, + 0.011273199692368507, + 0.0071976203471422195, + 0.007550384849309921, + 0.008858231827616692, + 0.009064107201993465, + 0.011217131279408932, + 0.008402070961892605, + 0.008561092428863049, + 0.00641328189522028, + 0.016724640503525734, + 0.01775827817618847, + 0.02913321740925312, + 0.016894256696105003, + 0.01875668577849865, + 0.013707264326512814, + 0.014936513267457485, + 0.015526908449828625, + 0.00866616889834404, + 0.006403605453670025, + 0.00999931339174509, + 0.004164583049714565, + 0.006162158213555813, + 0.005338622257113457, + 0.006398350931704044, + 0.005086577031761408, + 0.019345317035913467, + 0.03772522136569023, + 0.019430793821811676, + 0.017392925918102264, + 0.03274841606616974, + 0.02280941605567932, + 0.020194964483380318, + 0.030588174238801003, + 0.002037960337474942, + 0.003541793907061219, + 0.002309208968654275, + 0.0018691333243623376, + 0.0024521537125110626, + 0.0016038719331845641, + 0.0028322269208729267, + 0.0018144670175388455, + 0.006715684663504362, + 0.017870690673589706, + 0.00575864315032959, + 0.0033279939088970423, + 0.003995191305875778, + 0.007156034465879202, + 0.004006294999271631, + 0.00546282297000289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0067326370626688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle close-up captures a scene of profound tranquility and minimalist elegance. A vibrant crimson silk scarf, its fabric so light and sheer that its fine weave is almost translucent, lies spread across the smooth, dark mahogany surface of a polished wooden table. Its high-gloss finish reflects light like a still, dark pool of water. A gentle, unseen draft of air infuses the scene with energy, delicately lifting the scarf from the table. The fabric captures the current, first fluttering at the edge, then gracefully swelling upward in a slow-motion ballet of soft, tumbling waves. The silk rises and falls in a fluid, organic movement, then slowly, almost reluctantly, falls back down to drape the wood in a new, graceful form. The entire scene is bathed in a soft, directional light, as if from a window outside the frame, casting gentle, elongated shadows on the scarf and highlighting the silk's sheen against the deep, rich grain of the table, creating a striking, beautiful contrast.", + "seed": 27902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_0/index_7318_seed_27902_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": -0.05202238522469997, + "vlm_logprob": -0.05202238522469997, + "vlm_yes_no_margin": 3.174999952316284, + "vlm_token_credit": -0.05202238522469997, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047750528901815414 + ], + "opposite_logprob_tokens": [ + -3.422750473022461 + ], + "logprob_reward": -0.047750528901815414, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.047750528901815414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11327045410871506, + 0.044988926500082016, + 0.04556300491094589, + 0.030681652948260307, + 0.059618767350912094, + 0.14797809720039368, + 0.19244384765625, + 0.12222640216350555, + 0.20045556128025055, + 0.04277323931455612 + ], + "credit_map": [ + 0.011658367700874805, + 0.006719429977238178, + 0.015002932399511337, + 0.01908055506646633, + 0.013407660648226738, + 0.012132923118770123, + 0.01829216629266739, + 0.01697642356157303, + 0.0033411819022148848, + 0.004940856713801622, + 0.00503146555274725, + 0.006623533088713884, + 0.004970643203705549, + 0.004487909842282534, + 0.009497965686023235, + 0.006095373071730137, + 0.007095069158822298, + 0.0039839185774326324, + 0.006986223626881838, + 0.004752426873892546, + 0.004208291415125132, + 0.006798276212066412, + 0.005545170046389103, + 0.006193634122610092, + 0.002359403297305107, + 0.004237012006342411, + 0.0032766133081167936, + 0.0038229043129831553, + 0.004815211519598961, + 0.003578110598027706, + 0.004121648147702217, + 0.004470752086490393, + 0.004021130967885256, + 0.006098357029259205, + 0.007782269269227982, + 0.00805792585015297, + 0.00730816088616848, + 0.005062416661530733, + 0.010982302948832512, + 0.010306207463145256, + 0.009194177575409412, + 0.010859044268727303, + 0.01773083209991455, + 0.009189718402922153, + 0.04521804675459862, + 0.01052944641560316, + 0.012313415296375751, + 0.032943423837423325, + 0.010097889229655266, + 0.017802689224481583, + 0.015499814413487911, + 0.03608644753694534, + 0.027127321809530258, + 0.011552038602530956, + 0.03613942489027977, + 0.03813822194933891, + 0.006180324126034975, + 0.012723875232040882, + 0.020483309403061867, + 0.012851913459599018, + 0.013695474714040756, + 0.007031576242297888, + 0.02234587073326111, + 0.026914063841104507, + 0.011403881944715977, + 0.012109239585697651, + 0.019534600898623466, + 0.0172582920640707, + 0.023647991940379143, + 0.01962083764374256, + 0.05620546638965607, + 0.04067527502775192, + 0.0029815388843417168, + 0.002220084425061941, + 0.008671407587826252, + 0.0031520682387053967, + 0.0026209752541035414, + 0.004754639230668545, + 0.006786583457142115, + 0.011585947126150131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.047750528901815414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1130247637629509 + ], + "opposite_logprob_tokens": [ + -2.2380247116088867 + ], + "logprob_reward": -0.1130247637629509, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1130247637629509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14112967252731323, + 0.07580089569091797, + 0.04308141767978668, + 0.02595495618879795, + 0.040215443819761276, + 0.07804587483406067, + 0.09609970450401306, + 0.1782398521900177, + 0.27466443181037903, + 0.04676779359579086 + ], + "credit_map": [ + 0.015474444255232811, + 0.008668201044201851, + 0.02218693494796753, + 0.023233909159898758, + 0.01473077293485403, + 0.01463960763067007, + 0.021384185180068016, + 0.020811619237065315, + 0.00645593972876668, + 0.00402435939759016, + 0.012682955712080002, + 0.009429959580302238, + 0.005862066056579351, + 0.010002891533076763, + 0.010504897683858871, + 0.016837824136018753, + 0.0029816580936312675, + 0.004204058088362217, + 0.006571127101778984, + 0.004443333018571138, + 0.00519705330953002, + 0.0064568594098091125, + 0.006539494264870882, + 0.006687832530587912, + 0.001650750171393156, + 0.004465819336473942, + 0.0025990114081650972, + 0.00355212832801044, + 0.0044221519492566586, + 0.0021964549086987972, + 0.003821976948529482, + 0.0032466650009155273, + 0.003238300560042262, + 0.00362863065674901, + 0.006200532894581556, + 0.006842584349215031, + 0.005787787027657032, + 0.0038327446673065424, + 0.006008501630276442, + 0.004676364362239838, + 0.003917463589459658, + 0.005147124174982309, + 0.008781272917985916, + 0.004415005911141634, + 0.0201663076877594, + 0.004240871872752905, + 0.008191240951418877, + 0.023186588659882545, + 0.004942375235259533, + 0.009370354935526848, + 0.007946694269776344, + 0.012583618983626366, + 0.01477306243032217, + 0.0060488502494990826, + 0.017786292359232903, + 0.02264845184981823, + 0.013241875916719437, + 0.02259407751262188, + 0.01890963688492775, + 0.019078953191637993, + 0.02100669965147972, + 0.013534673489630222, + 0.03183850646018982, + 0.03803543373942375, + 0.016494888812303543, + 0.01609189808368683, + 0.029059993103146553, + 0.02432580664753914, + 0.028718581423163414, + 0.0296822227537632, + 0.07571974396705627, + 0.05457128584384918, + 0.003174689132720232, + 0.0030121186282485723, + 0.005550675094127655, + 0.004271316807717085, + 0.0038735733833163977, + 0.005621138494461775, + 0.009301135316491127, + 0.01196314673870802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.1130247637629509 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809603676199913 + ], + "opposite_logprob_tokens": [ + -3.2880959510803223 + ], + "logprob_reward": -0.03809603676199913, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809603676199913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10524352639913559, + 0.08587758243083954, + 0.05597608909010887, + 0.037506990134716034, + 0.15153706073760986, + 0.17164354026317596, + 0.11825314164161682, + 0.07087992131710052, + 0.1361331343650818, + 0.0669490322470665 + ], + "credit_map": [ + 0.013886476866900921, + 0.007060926873236895, + 0.013785301707684994, + 0.017348533496260643, + 0.011323707178235054, + 0.01153107825666666, + 0.016366690397262573, + 0.013940811157226562, + 0.009339881129562855, + 0.0043006762862205505, + 0.014701117761433125, + 0.010385955683887005, + 0.00547871133312583, + 0.01008769404143095, + 0.010973016731441021, + 0.02061053365468979, + 0.00605337368324399, + 0.004059756174683571, + 0.01149697881191969, + 0.007247613742947578, + 0.004954625386744738, + 0.0073874881491065025, + 0.007323297206312418, + 0.0074529582634568214, + 0.004054160788655281, + 0.0032399764750152826, + 0.006504547316581011, + 0.003503781510517001, + 0.004241385031491518, + 0.005665595643222332, + 0.004466511774808168, + 0.00583103159442544, + 0.010748583823442459, + 0.009242715314030647, + 0.025884520262479782, + 0.022413328289985657, + 0.0190779659897089, + 0.01387881487607956, + 0.019878080114722252, + 0.030413052067160606, + 0.017973052337765694, + 0.010264184325933456, + 0.040267281234264374, + 0.019466524943709373, + 0.015440152026712894, + 0.014033895917236805, + 0.02296566031873226, + 0.03123280592262745, + 0.008176857605576515, + 0.01028952095657587, + 0.016423184424638748, + 0.012739897705614567, + 0.018171600997447968, + 0.008832701481878757, + 0.020710889250040054, + 0.022908488288521767, + 0.004273355007171631, + 0.00588100403547287, + 0.010049596428871155, + 0.006851469166576862, + 0.005803682841360569, + 0.004218106623739004, + 0.015766000375151634, + 0.018036706373095512, + 0.008107172325253487, + 0.006479957606643438, + 0.014276053756475449, + 0.012138871476054192, + 0.013011695817112923, + 0.00972031895071268, + 0.04215983673930168, + 0.030239224433898926, + 0.00389577797614038, + 0.0021361829712986946, + 0.013357270509004593, + 0.004240021109580994, + 0.0035907109268009663, + 0.004739224445074797, + 0.013468369841575623, + 0.021521475166082382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.03809603676199913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305575042963028 + ], + "opposite_logprob_tokens": [ + -3.168055772781372 + ], + "logprob_reward": -0.04305575042963028, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305575042963028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1571449488401413, + 0.035511985421180725, + 0.029826659709215164, + 0.0161878764629364, + 0.049810659140348434, + 0.07834077626466751, + 0.1398223489522934, + 0.11331167817115784, + 0.12523239850997925, + 0.2548106908798218 + ], + "credit_map": [ + 0.01739673502743244, + 0.009528597816824913, + 0.02099829725921154, + 0.021126238629221916, + 0.018984731286764145, + 0.015935298055410385, + 0.023556146770715714, + 0.02961891144514084, + 0.00241638720035553, + 0.0028608026914298534, + 0.00445671658962965, + 0.004598539788275957, + 0.004956537391990423, + 0.005681318696588278, + 0.005463093053549528, + 0.005078587215393782, + 0.002256829757243395, + 0.00342272175475955, + 0.0038909518625587225, + 0.004797790199518204, + 0.0037788893096148968, + 0.0030963399913161993, + 0.004478753544390202, + 0.004104381427168846, + 0.0019679907709360123, + 0.0017337214667350054, + 0.0018319588853046298, + 0.0016814826522022486, + 0.0021057541016489267, + 0.0017162543954327703, + 0.002862254623323679, + 0.0022884595673531294, + 0.004042278975248337, + 0.0064650108106434345, + 0.0070977709256112576, + 0.007485183421522379, + 0.00725113321095705, + 0.005638797767460346, + 0.005799493752419949, + 0.006030992139130831, + 0.005149233154952526, + 0.009187036193907261, + 0.009978457354009151, + 0.008817977271974087, + 0.017375878989696503, + 0.004942398518323898, + 0.010199757292866707, + 0.01269003376364708, + 0.006845036055892706, + 0.010090643540024757, + 0.021225741133093834, + 0.014948825351893902, + 0.013486658222973347, + 0.00799570232629776, + 0.024992497637867928, + 0.040237244218587875, + 0.006652677431702614, + 0.008541015908122063, + 0.023424513638019562, + 0.007185208611190319, + 0.00869904737919569, + 0.006223853211849928, + 0.019195953384041786, + 0.033389415591955185, + 0.0075232102535665035, + 0.006927802227437496, + 0.012478862889111042, + 0.01076533179730177, + 0.012396425008773804, + 0.015505289658904076, + 0.03447077050805092, + 0.025164702907204628, + 0.013272377662360668, + 0.006972215604037046, + 0.04781004786491394, + 0.015290376730263233, + 0.011675778776407242, + 0.014269353821873665, + 0.056694090366363525, + 0.08882643282413483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.04305575042963028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vibrant crimson silk scarf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scarf's fabric light, sheer, and translucent enough to show its fine weave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scarf lying on a smooth, dark mahogany wooden table with a high-gloss finish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scarf move in a slow-motion sequence of fluttering, swelling upward in waves, and then falling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft, directional light casting gentle, elongated shadows and highlighting the silk's sheen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818484626710415 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.01818484626710415, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818484626710415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18459893763065338, + 0.04858430102467537, + 0.033113326877355576, + 0.02031761407852173, + 0.028084034100174904, + 0.1131787896156311, + 0.08663967251777649, + 0.25714945793151855, + 0.14518778026103973, + 0.0831461101770401 + ], + "credit_map": [ + 0.021133925765752792, + 0.01317967101931572, + 0.02915003150701523, + 0.029899802058935165, + 0.021126309409737587, + 0.01929311640560627, + 0.02643970213830471, + 0.02437635324895382, + 0.004552942235022783, + 0.0030155170243233442, + 0.007484366651624441, + 0.0065746731124818325, + 0.00403140764683485, + 0.005236586555838585, + 0.007107040844857693, + 0.010581761598587036, + 0.004431396257132292, + 0.0030884251464158297, + 0.00590924359858036, + 0.0032053187023848295, + 0.0028642641846090555, + 0.004960370250046253, + 0.0034749603364616632, + 0.005179346539080143, + 0.0016829256201162934, + 0.0019503823714330792, + 0.004163878504186869, + 0.0024322334211319685, + 0.0024621128104627132, + 0.0022975991014391184, + 0.002020658925175667, + 0.00330782076343894, + 0.002202758099883795, + 0.0025737553369253874, + 0.004132784437388182, + 0.004916954319924116, + 0.003920187242329121, + 0.0024786563590168953, + 0.004231192637234926, + 0.003627741476520896, + 0.007004555314779282, + 0.008702038787305355, + 0.018675832077860832, + 0.009661958552896976, + 0.031444016844034195, + 0.007643628399819136, + 0.009558344259858131, + 0.020488407462835312, + 0.005882766097784042, + 0.008536204695701599, + 0.011139583773911, + 0.00971259269863367, + 0.013301973231136799, + 0.0071367379277944565, + 0.01440425869077444, + 0.016525547951459885, + 0.012608052231371403, + 0.014242054894566536, + 0.053279902786016464, + 0.017477406188845634, + 0.012789176777005196, + 0.013183494098484516, + 0.052045680582523346, + 0.08152367174625397, + 0.008564690127968788, + 0.0076380642130970955, + 0.015371391549706459, + 0.013848225586116314, + 0.014113391749560833, + 0.014985712245106697, + 0.04113588109612465, + 0.029530411586165428, + 0.004288098309189081, + 0.0024592611007392406, + 0.016632461920380592, + 0.005066426936537027, + 0.003271970897912979, + 0.005363871343433857, + 0.017885545268654823, + 0.028178464621305466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.01818484626710415 + } + ], + "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_1/index_tensor([176, 176])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_tensor([176, 176])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5ef43106878bfee4f114a2a3b2a043ce9f912761 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_tensor([176, 176])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, handheld close-up captures the scene with visceral immediacy, the camera shaking dramatically in perfect sync with the action. A man's powerful hands dominate the frame; one grips the neck of an icy, brown glass beer bottle, smeared with condensation, while the other decisively hooks a simple stainless steel bottle opener beneath the metal cap. With a swift upward flick of the wrist, the cap pops free, and in that moment, the camera shakes, mimicking the force of the impact. Then, the man's hand casually drops the opener, its sound just out of focus, landing on the warm-toned wooden surface in the background. Without hesitation, he steadily pushes the bottle's mouth toward the camera, tilts it back, and takes a long, satisfying sip. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long shadows and creating shimmers on the damp glass, evoking a relaxed, lighthearted atmosphere.", + "seed": 28407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_176_seed_28407_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": -0.07042035423219203, + "vlm_logprob": -0.07042035423219203, + "vlm_yes_no_margin": 3.099999952316284, + "vlm_token_credit": -0.07042035423219203, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.022361863404512405 + ], + "opposite_logprob_tokens": [ + -4.647361755371094 + ], + "logprob_reward": -0.022361863404512405, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.022361863404512405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15460580587387085, + 0.10518858581781387, + 0.07023509591817856, + 0.1451578587293625, + 0.12233543395996094, + 0.15433275699615479, + 0.06300786137580872, + 0.04690813273191452, + 0.080194391310215, + 0.058034081012010574 + ], + "credit_map": [ + 0.011654150672256947, + 0.017618825659155846, + 0.02418125793337822, + 0.03484367951750755, + 0.01663678139448166, + 0.02331702597439289, + 0.01050659827888012, + 0.01584748737514019, + 0.00946867372840643, + 0.010250547900795937, + 0.01340683363378048, + 0.013894309289753437, + 0.01577823795378208, + 0.009882868267595768, + 0.020913276821374893, + 0.0115938326343894, + 0.0061792596243321896, + 0.011005361564457417, + 0.012842939235270023, + 0.008413303643465042, + 0.007604220416396856, + 0.005701358430087566, + 0.011546454392373562, + 0.006942199543118477, + 0.009254220873117447, + 0.01666194759309292, + 0.021121898666024208, + 0.025436578318476677, + 0.03640822321176529, + 0.00966403167694807, + 0.013748283497989178, + 0.01286267675459385, + 0.009226190857589245, + 0.016020692884922028, + 0.014904404990375042, + 0.012043609283864498, + 0.01932980678975582, + 0.014326125383377075, + 0.021137168630957603, + 0.015347431413829327, + 0.014962142333388329, + 0.018050922080874443, + 0.029323922470211983, + 0.03300965204834938, + 0.014807075262069702, + 0.013210587203502655, + 0.008328764699399471, + 0.022639691829681396, + 0.019369101151823997, + 0.009161756373941898, + 0.006548992823809385, + 0.005401125177741051, + 0.00694451667368412, + 0.0047845239751040936, + 0.004799391143023968, + 0.005998446140438318, + 0.008301524445414543, + 0.004578292835503817, + 0.006862940266728401, + 0.0060389987193048, + 0.005028223618865013, + 0.004934143740683794, + 0.00613348139449954, + 0.005030526779592037, + 0.009987684898078442, + 0.008245931006968021, + 0.008420434780418873, + 0.009563678875565529, + 0.013442008756101131, + 0.009172482416033745, + 0.012441135942935944, + 0.008921035565435886, + 0.0067815398797392845, + 0.00851368997246027, + 0.010066402144730091, + 0.005327840335667133, + 0.009760219603776932, + 0.007061904761940241, + 0.004181834403425455, + 0.006340651772916317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.022361863404512405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.018223943188786507 + ], + "opposite_logprob_tokens": [ + -4.018223762512207 + ], + "logprob_reward": -0.018223943188786507, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018223943188786507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1162407323718071, + 0.32778868079185486, + 0.11038319021463394, + 0.06581345200538635, + 0.07476785033941269, + 0.05636139586567879, + 0.04533442109823227, + 0.09587671607732773, + 0.08284419029951096, + 0.024589378386735916 + ], + "credit_map": [ + 0.010420206934213638, + 0.008610676042735577, + 0.015325089916586876, + 0.019932443276047707, + 0.016535185277462006, + 0.012569224461913109, + 0.018849946558475494, + 0.013997961767017841, + 0.03132743760943413, + 0.03393017128109932, + 0.043673593550920486, + 0.0379471555352211, + 0.03866050764918327, + 0.060346126556396484, + 0.04703092947602272, + 0.03487275913357735, + 0.014109066687524319, + 0.013861478306353092, + 0.013814052566885948, + 0.011241663247346878, + 0.017471102997660637, + 0.010313262231647968, + 0.017633022740483284, + 0.011939543299376965, + 0.0038217692635953426, + 0.009728260338306427, + 0.009483659639954567, + 0.009727207012474537, + 0.018158236518502235, + 0.004084945656359196, + 0.005665693432092667, + 0.005143688526004553, + 0.005766707472503185, + 0.013694442808628082, + 0.009769859723746777, + 0.007679393980652094, + 0.014388136565685272, + 0.005875617731362581, + 0.009721643291413784, + 0.00787204597145319, + 0.0050379266031086445, + 0.006042941473424435, + 0.009972810745239258, + 0.00946517288684845, + 0.006066541187465191, + 0.0061011407524347305, + 0.007194094359874725, + 0.006480770651251078, + 0.005706681404262781, + 0.010692783631384373, + 0.0037644144613295794, + 0.004126749932765961, + 0.003798009594902396, + 0.003943300805985928, + 0.002899659564718604, + 0.01040282379835844, + 0.010808193124830723, + 0.0064126490615308285, + 0.012457245029509068, + 0.01135177444666624, + 0.012158016674220562, + 0.014866946265101433, + 0.014751830138266087, + 0.01307005900889635, + 0.006554721388965845, + 0.006876024883240461, + 0.012686194851994514, + 0.015310180373489857, + 0.008535937406122684, + 0.00912789162248373, + 0.014331396669149399, + 0.009421841241419315, + 0.0022386668715626, + 0.003197664627805352, + 0.002470768056809902, + 0.0024055689573287964, + 0.00442836107686162, + 0.0025695955846458673, + 0.004365880973637104, + 0.0029128724709153175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.018223943188786507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.062096573412418365 + ], + "opposite_logprob_tokens": [ + -2.81209659576416 + ], + "logprob_reward": -0.062096573412418365, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062096573412418365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09736239910125732, + 0.18059588968753815, + 0.15280970931053162, + 0.15343236923217773, + 0.026266617700457573, + 0.1145024299621582, + 0.0913589671254158, + 0.059151552617549896, + 0.061742257326841354, + 0.06277777999639511 + ], + "credit_map": [ + 0.009212914854288101, + 0.009830120950937271, + 0.015430835075676441, + 0.014412693679332733, + 0.009008371271193027, + 0.01675664633512497, + 0.00964528601616621, + 0.013065527193248272, + 0.015203375369310379, + 0.02279885858297348, + 0.026890061795711517, + 0.02505011297762394, + 0.020068859681487083, + 0.02380908839404583, + 0.027876297011971474, + 0.01889924146234989, + 0.010290262289345264, + 0.019505802541971207, + 0.022620422765612602, + 0.016836119815707207, + 0.01963716745376587, + 0.011803870089352131, + 0.039167724549770355, + 0.012948346324265003, + 0.009273550473153591, + 0.02276306040585041, + 0.02153918333351612, + 0.02336234226822853, + 0.04050818085670471, + 0.008944029919803143, + 0.013168426230549812, + 0.013873595744371414, + 0.0018067221390083432, + 0.0051605491898953915, + 0.0032334979623556137, + 0.003341173753142357, + 0.00516184838488698, + 0.0016220748657360673, + 0.00306192715652287, + 0.002878825645893812, + 0.008769754320383072, + 0.010598578490316868, + 0.017371851950883865, + 0.023530911654233932, + 0.012883329764008522, + 0.010298644192516804, + 0.013428117148578167, + 0.0176212340593338, + 0.00849122740328312, + 0.03055196814239025, + 0.008357449434697628, + 0.005623819772154093, + 0.006192285101860762, + 0.007308390457183123, + 0.004145499784499407, + 0.020688330754637718, + 0.007527271285653114, + 0.004500577691942453, + 0.0072830962017178535, + 0.006205165758728981, + 0.00869949534535408, + 0.0051093329675495625, + 0.011989081278443336, + 0.00783753301948309, + 0.004200399853289127, + 0.005032030399888754, + 0.00870442483574152, + 0.010292201302945614, + 0.007330083288252354, + 0.00712518161162734, + 0.011433643288910389, + 0.007624294143170118, + 0.004735586699098349, + 0.009698661975562572, + 0.006381802726536989, + 0.0061206393875181675, + 0.016133487224578857, + 0.005465012043714523, + 0.008258487097918987, + 0.005984096322208643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.062096573412418365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.16037480533123016 + ], + "opposite_logprob_tokens": [ + -1.9103747606277466 + ], + "logprob_reward": -0.16037480533123016, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037480533123016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10572125017642975, + 0.09304124861955643, + 0.07320750504732132, + 0.16032451391220093, + 0.07918297499418259, + 0.17033730447292328, + 0.06408042460680008, + 0.0886165127158165, + 0.1013316661119461, + 0.06415659189224243 + ], + "credit_map": [ + 0.010745863430202007, + 0.00751393660902977, + 0.01699822209775448, + 0.019500115886330605, + 0.012567970901727676, + 0.013715362176299095, + 0.011967121623456478, + 0.012712654657661915, + 0.008981327526271343, + 0.009129942394793034, + 0.013006221503019333, + 0.012520913034677505, + 0.010450907982885838, + 0.012939308770000935, + 0.014253087341785431, + 0.01175953634083271, + 0.006607119459658861, + 0.011377228423953056, + 0.010430141352117062, + 0.007356756832450628, + 0.009755336679518223, + 0.008058062754571438, + 0.011617183685302734, + 0.008005679585039616, + 0.009697218425571918, + 0.021054523065686226, + 0.0233943872153759, + 0.0231401976197958, + 0.04445421323180199, + 0.009638305753469467, + 0.013488619588315487, + 0.01545704621821642, + 0.00577491195872426, + 0.008637815713882446, + 0.01428403239697218, + 0.01192973367869854, + 0.011714663356542587, + 0.006156259216368198, + 0.011501181870698929, + 0.009184370748698711, + 0.01816878281533718, + 0.018186217173933983, + 0.023171376436948776, + 0.020865274593234062, + 0.027488447725772858, + 0.027847623452544212, + 0.014233683235943317, + 0.020375898107886314, + 0.008072284050285816, + 0.012323897331953049, + 0.005868210457265377, + 0.0038461266085505486, + 0.006623241119086742, + 0.007399075664579868, + 0.006112057715654373, + 0.013835534453392029, + 0.008975710719823837, + 0.0070128957740962505, + 0.012158825062215328, + 0.010349997319281101, + 0.011530444957315922, + 0.013002586551010609, + 0.014233195222914219, + 0.011352854780852795, + 0.007165186107158661, + 0.007518269121646881, + 0.015118148177862167, + 0.019698843359947205, + 0.010982459411025047, + 0.011676346883177757, + 0.01818563975393772, + 0.010986775159835815, + 0.00633629085496068, + 0.0099833058193326, + 0.007471611723303795, + 0.006155029404908419, + 0.014986026100814342, + 0.004736605565994978, + 0.007897408679127693, + 0.0065903132781386375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.16037480533123016 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.08904458582401276 + ], + "opposite_logprob_tokens": [ + -2.4640445709228516 + ], + "logprob_reward": -0.08904458582401276, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904458582401276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11777766793966293, + 0.08244850486516953, + 0.16120561957359314, + 0.025937292724847794, + 0.056462083011865616, + 0.06494428962469101, + 0.24162308871746063, + 0.08986078947782516, + 0.14171075820922852, + 0.01802990585565567 + ], + "credit_map": [ + 0.015314677730202675, + 0.008654039353132248, + 0.01706971600651741, + 0.024636726826429367, + 0.012895530089735985, + 0.010368023067712784, + 0.015203491784632206, + 0.013635463081300259, + 0.008128519169986248, + 0.009395022876560688, + 0.012145277112722397, + 0.011270292103290558, + 0.008374364115297794, + 0.010859190486371517, + 0.01165894977748394, + 0.010616879910230637, + 0.017735302448272705, + 0.02654855325818062, + 0.025245632976293564, + 0.019738251343369484, + 0.01868261583149433, + 0.016038643196225166, + 0.018323106691241264, + 0.018893534317612648, + 0.0021728896535933018, + 0.0036142668686807156, + 0.0034604251850396395, + 0.004379532765597105, + 0.004694790579378605, + 0.002499779686331749, + 0.00252702203579247, + 0.002588583854958415, + 0.004749809857457876, + 0.010443328879773617, + 0.006631304509937763, + 0.006098940037190914, + 0.009432743303477764, + 0.006448501255363226, + 0.006942931097000837, + 0.00571452546864748, + 0.0060152155347168446, + 0.007854516617953777, + 0.012905010022222996, + 0.015253896825015545, + 0.0059382678009569645, + 0.005498562008142471, + 0.004078800790011883, + 0.007400013972073793, + 0.02096289023756981, + 0.08850903064012527, + 0.01204659603536129, + 0.012254172936081886, + 0.010361598804593086, + 0.011314515955746174, + 0.007166340947151184, + 0.07900793105363846, + 0.010935017839074135, + 0.006653313525021076, + 0.012463939376175404, + 0.010448221117258072, + 0.00981387123465538, + 0.01196617353707552, + 0.014852285385131836, + 0.012727961875498295, + 0.0185100045055151, + 0.016611559316515923, + 0.020329367369413376, + 0.027981238439679146, + 0.012992819771170616, + 0.012472008354961872, + 0.017884623259305954, + 0.014929141849279404, + 0.0017830251017585397, + 0.0028796193655580282, + 0.003383628325536847, + 0.0019810956437140703, + 0.0022086065728217363, + 0.0015315408818423748, + 0.002415468916296959, + 0.0018469208152964711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.08904458582401276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, handheld close-up captures the scene with visceral immediacy, the camera shaking dramatically in perfect sync with the action. A man's powerful hands dominate the frame; one grips the neck of an icy, brown glass beer bottle, smeared with condensation, while the other decisively hooks a simple stainless steel bottle opener beneath the metal cap. With a swift upward flick of the wrist, the cap pops free, and in that moment, the camera shakes, mimicking the force of the impact. Then, the man's hand casually drops the opener, its sound just out of focus, landing on the warm-toned wooden surface in the background. Without hesitation, he steadily pushes the bottle's mouth toward the camera, tilts it back, and takes a long, satisfying sip. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long shadows and creating shimmers on the damp glass, evoking a relaxed, lighthearted atmosphere.", + "seed": 28408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_176_seed_28408_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": -0.06502971816807986, + "vlm_logprob": -0.06502971816807986, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.06502971816807986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.022046716883778572 + ], + "opposite_logprob_tokens": [ + -5.1470465660095215 + ], + "logprob_reward": -0.022046716883778572, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.022046716883778572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09874536097049713, + 0.05250261351466179, + 0.11927685886621475, + 0.05064276233315468, + 0.27738431096076965, + 0.06909237802028656, + 0.04917607828974724, + 0.04201365262269974, + 0.10490911453962326, + 0.1362568736076355 + ], + "credit_map": [ + 0.014967207796871662, + 0.010033722035586834, + 0.007577070035040379, + 0.013925920240581036, + 0.006746428087353706, + 0.013916634023189545, + 0.0169678907841444, + 0.014610488899052143, + 0.004121924750506878, + 0.00745596457272768, + 0.0036176100838929415, + 0.005870271008461714, + 0.009069372899830341, + 0.0058157360181212425, + 0.009246647357940674, + 0.007305085193365812, + 0.009572473354637623, + 0.020648250356316566, + 0.010429080575704575, + 0.02264511212706566, + 0.013410951010882854, + 0.010907668620347977, + 0.0174445528537035, + 0.014218772761523724, + 0.006431834306567907, + 0.003974701277911663, + 0.0037764315493404865, + 0.00928958784788847, + 0.0038488635327667, + 0.0073356181383132935, + 0.008488941006362438, + 0.007496784441173077, + 0.01930117793381214, + 0.040506601333618164, + 0.024712594226002693, + 0.03220810368657112, + 0.02576017752289772, + 0.051502883434295654, + 0.030539648607373238, + 0.052853140980005264, + 0.007737630046904087, + 0.007112281396985054, + 0.009569027461111546, + 0.008535155095160007, + 0.003365211421623826, + 0.006198088172823191, + 0.012653940357267857, + 0.013921044766902924, + 0.003964120987802744, + 0.0037666969001293182, + 0.0036013617645949125, + 0.0048350621946156025, + 0.006727958098053932, + 0.006337674334645271, + 0.013180249370634556, + 0.0067629567347466946, + 0.004909280687570572, + 0.005714915692806244, + 0.003899322124198079, + 0.005687021184712648, + 0.005114613566547632, + 0.005346501711755991, + 0.006517224945127964, + 0.0048247696831822395, + 0.014648323878645897, + 0.014813128858804703, + 0.008288666605949402, + 0.010272138752043247, + 0.014571050181984901, + 0.013306805863976479, + 0.012413082644343376, + 0.016595911234617233, + 0.015601721592247486, + 0.015674808993935585, + 0.022033022716641426, + 0.02361368201673031, + 0.015386899001896381, + 0.014184484258294106, + 0.018330460414290428, + 0.011431792750954628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.022046716883778572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.01256749127060175 + ], + "opposite_logprob_tokens": [ + -4.387567520141602 + ], + "logprob_reward": -0.01256749127060175, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256749127060175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1313561350107193, + 0.14699774980545044, + 0.11557214707136154, + 0.10095823556184769, + 0.15453943610191345, + 0.1655767858028412, + 0.0878930389881134, + 0.02098236419260502, + 0.0343756340444088, + 0.0417485237121582 + ], + "credit_map": [ + 0.012480407021939754, + 0.011851304210722446, + 0.010561824776232243, + 0.018763476982712746, + 0.01673056371510029, + 0.012476710602641106, + 0.027814624831080437, + 0.02067721262574196, + 0.010253362357616425, + 0.01720389351248741, + 0.011190026998519897, + 0.033827558159828186, + 0.018712084740400314, + 0.01693158596754074, + 0.020935416221618652, + 0.01794380694627762, + 0.008905377238988876, + 0.01746838539838791, + 0.013536443933844566, + 0.01644486002624035, + 0.01617453247308731, + 0.01172042265534401, + 0.015554212965071201, + 0.0157678984105587, + 0.007007727399468422, + 0.006098340731114149, + 0.02127011865377426, + 0.010532866232097149, + 0.0051029217429459095, + 0.009883655235171318, + 0.010544218122959137, + 0.03051837719976902, + 0.011314118281006813, + 0.016830040141940117, + 0.013882162049412727, + 0.026081867516040802, + 0.010690783150494099, + 0.020457522943615913, + 0.03504364565014839, + 0.020239299163222313, + 0.03498362749814987, + 0.016333043575286865, + 0.023673828691244125, + 0.017573243007063866, + 0.006864955648779869, + 0.018221627920866013, + 0.01231659296900034, + 0.035609837621450424, + 0.005940149072557688, + 0.01059640571475029, + 0.0062652169726789, + 0.007777644321322441, + 0.015620938502252102, + 0.013449138030409813, + 0.01837787590920925, + 0.00986566487699747, + 0.0015302159590646625, + 0.0018721643136814237, + 0.002029963070526719, + 0.002745729172602296, + 0.0028586192056536674, + 0.0027458774857223034, + 0.004910557996481657, + 0.002289237454533577, + 0.005015903152525425, + 0.0038548342417925596, + 0.00259006442502141, + 0.003342675045132637, + 0.0051722717471420765, + 0.003382213180884719, + 0.004764708224684, + 0.0062529644928872585, + 0.003923229873180389, + 0.0051532285287976265, + 0.0064672562293708324, + 0.004425723105669022, + 0.0063180578872561455, + 0.005834400653839111, + 0.006010835058987141, + 0.003615789348259568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.01256749127060175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.029818108305335045 + ], + "opposite_logprob_tokens": [ + -3.529818058013916 + ], + "logprob_reward": -0.029818108305335045, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029818108305335045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09008878469467163, + 0.0348970964550972, + 0.1768331676721573, + 0.18570436537265778, + 0.06333328783512115, + 0.3285369873046875, + 0.05009559169411659, + 0.017940737307071686, + 0.01857762411236763, + 0.03399239107966423 + ], + "credit_map": [ + 0.007689703721553087, + 0.008206670172512531, + 0.007012110203504562, + 0.014143235050141811, + 0.013137219473719597, + 0.008540532551705837, + 0.017489582300186157, + 0.013869731687009335, + 0.0026318137533962727, + 0.00516885332763195, + 0.002550101839005947, + 0.005615403410047293, + 0.005376447457820177, + 0.004430024418979883, + 0.004139898344874382, + 0.004984552040696144, + 0.014385782182216644, + 0.027600953355431557, + 0.01809529960155487, + 0.02219000644981861, + 0.023180847987532616, + 0.018183482810854912, + 0.020016497001051903, + 0.03318029269576073, + 0.012082912027835846, + 0.005171927623450756, + 0.053311511874198914, + 0.011102425865828991, + 0.00483535835519433, + 0.016012834385037422, + 0.009036281146109104, + 0.07415110617876053, + 0.006114413496106863, + 0.0061396146193146706, + 0.007409927435219288, + 0.009172460064291954, + 0.004869102034717798, + 0.009192352183163166, + 0.00943332351744175, + 0.011002094484865665, + 0.05593084916472435, + 0.032304547727108, + 0.04239339753985405, + 0.038638465106487274, + 0.030086863785982132, + 0.026233112439513206, + 0.05071870982646942, + 0.05223105475306511, + 0.0035645603202283382, + 0.004295608960092068, + 0.003333538770675659, + 0.004689008928835392, + 0.007222771644592285, + 0.0066294376738369465, + 0.014044033363461494, + 0.006316630635410547, + 0.0015199951594695449, + 0.001984583679586649, + 0.001638885005377233, + 0.0016473096329718828, + 0.003055656561627984, + 0.002384164137765765, + 0.003469438524916768, + 0.0022407048381865025, + 0.002317160600796342, + 0.002015200676396489, + 0.001627015764825046, + 0.0013688921462744474, + 0.003024236299097538, + 0.002080774400383234, + 0.0024101415183395147, + 0.0037342002615332603, + 0.0035517076030373573, + 0.0038902987726032734, + 0.003006202867254615, + 0.005019770003855228, + 0.006283883471041918, + 0.004469880368560553, + 0.004559226334095001, + 0.0032114239875227213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.029818108305335045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.16038718819618225 + ], + "opposite_logprob_tokens": [ + -1.9103871583938599 + ], + "logprob_reward": -0.16038718819618225, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038718819618225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04817303270101547, + 0.08243151009082794, + 0.08146438747644424, + 0.1860395073890686, + 0.2001701295375824, + 0.054673414677381516, + 0.10908984392881393, + 0.10130714625120163, + 0.08241122961044312, + 0.054239850491285324 + ], + "credit_map": [ + 0.0055642444640398026, + 0.004750823136419058, + 0.0038449731655418873, + 0.006453008856624365, + 0.005059362389147282, + 0.00723938038572669, + 0.008266259916126728, + 0.006994980853050947, + 0.015986653044819832, + 0.008761819452047348, + 0.006420336663722992, + 0.00903573352843523, + 0.008447065949440002, + 0.006960698403418064, + 0.008167056366801262, + 0.01865214668214321, + 0.006973113864660263, + 0.010979664511978626, + 0.006522071082144976, + 0.010885704308748245, + 0.013537263497710228, + 0.008394226431846619, + 0.012086180038750172, + 0.012086154893040657, + 0.014622495509684086, + 0.009830047376453876, + 0.039910655468702316, + 0.01704993098974228, + 0.008478553034365177, + 0.019381476566195488, + 0.019276972860097885, + 0.05748938396573067, + 0.01667601987719536, + 0.026906339451670647, + 0.019786030054092407, + 0.02764110267162323, + 0.014947501942515373, + 0.03536001592874527, + 0.02471861056983471, + 0.034134503453969955, + 0.010213339701294899, + 0.005398879759013653, + 0.00706780981272459, + 0.00595862278714776, + 0.0035532349720597267, + 0.0070323338732123375, + 0.0060439337976276875, + 0.009405259042978287, + 0.008300406858325005, + 0.017228703945875168, + 0.008644617162644863, + 0.011694595217704773, + 0.015118610113859177, + 0.01665380783379078, + 0.016983719542622566, + 0.014465381391346455, + 0.015639787539839745, + 0.014864944852888584, + 0.010133431293070316, + 0.011790817603468895, + 0.012638280168175697, + 0.012842775322496891, + 0.014074429869651794, + 0.00932268239557743, + 0.015860995277762413, + 0.011728564277291298, + 0.007102704606950283, + 0.007101831026375294, + 0.00936864223331213, + 0.011529413983225822, + 0.008485752157866955, + 0.011233322322368622, + 0.0067866044119000435, + 0.0062134964391589165, + 0.007008627522736788, + 0.006750199943780899, + 0.006954746786504984, + 0.007346350699663162, + 0.005992009770125151, + 0.007187813054770231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.16038718819618225 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.10032908618450165 + ], + "opposite_logprob_tokens": [ + -2.3503291606903076 + ], + "logprob_reward": -0.10032908618450165, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032908618450165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1863103210926056, + 0.12734727561473846, + 0.050060492008924484, + 0.2294233739376068, + 0.03403346240520477, + 0.0753004401922226, + 0.1212925910949707, + 0.05936074256896973, + 0.06797612458467484, + 0.048895228654146194 + ], + "credit_map": [ + 0.03913479298353195, + 0.04002859443426132, + 0.015114295296370983, + 0.024345695972442627, + 0.014326244592666626, + 0.015439375303685665, + 0.01776769384741783, + 0.020153608173131943, + 0.021578244864940643, + 0.01215858943760395, + 0.00879783183336258, + 0.019092150032520294, + 0.009314583614468575, + 0.012444984167814255, + 0.013807554729282856, + 0.03015332855284214, + 0.004366403445601463, + 0.005610911175608635, + 0.00624713534489274, + 0.008329027332365513, + 0.005981288384646177, + 0.004847323056310415, + 0.005893881898373365, + 0.00878452230244875, + 0.01584460586309433, + 0.0073426468297839165, + 0.06276306509971619, + 0.013919226825237274, + 0.0059021711349487305, + 0.022952331230044365, + 0.01386252697557211, + 0.0868367925286293, + 0.002667022403329611, + 0.0024480128195136786, + 0.002190516795963049, + 0.007274616975337267, + 0.0019300011917948723, + 0.005740117747336626, + 0.006910842843353748, + 0.004872325342148542, + 0.014825900085270405, + 0.007560740225017071, + 0.010266421362757683, + 0.007285286206752062, + 0.0032795227598398924, + 0.008478152565658092, + 0.007952697575092316, + 0.015651708468794823, + 0.013408152386546135, + 0.012557575479149818, + 0.011506752111017704, + 0.012706847861409187, + 0.01630244590342045, + 0.016724606975913048, + 0.019325701519846916, + 0.018760504201054573, + 0.01027995627373457, + 0.010178063064813614, + 0.006145525258034468, + 0.00802850816398859, + 0.006243497598916292, + 0.004808097146451473, + 0.007336033508181572, + 0.006341052241623402, + 0.007644441910088062, + 0.010337148793041706, + 0.00586656341329217, + 0.00906434003263712, + 0.00938095897436142, + 0.00957284402102232, + 0.008006218820810318, + 0.008103604428470135, + 0.005211463198065758, + 0.00582427904009819, + 0.00894190650433302, + 0.0060988981276750565, + 0.006465766578912735, + 0.006459415890276432, + 0.006208368577063084, + 0.003685125382617116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.10032908618450165 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, handheld close-up captures the scene with visceral immediacy, the camera shaking dramatically in perfect sync with the action. A man's powerful hands dominate the frame; one grips the neck of an icy, brown glass beer bottle, smeared with condensation, while the other decisively hooks a simple stainless steel bottle opener beneath the metal cap. With a swift upward flick of the wrist, the cap pops free, and in that moment, the camera shakes, mimicking the force of the impact. Then, the man's hand casually drops the opener, its sound just out of focus, landing on the warm-toned wooden surface in the background. Without hesitation, he steadily pushes the bottle's mouth toward the camera, tilts it back, and takes a long, satisfying sip. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long shadows and creating shimmers on the damp glass, evoking a relaxed, lighthearted atmosphere.", + "seed": 28409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_176_seed_28409_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": -0.10039787590503693, + "vlm_logprob": -0.10039787590503693, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.10039787590503693, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.020198959857225418 + ], + "opposite_logprob_tokens": [ + -4.395198822021484 + ], + "logprob_reward": -0.020198959857225418, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.020198959857225418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2007025182247162, + 0.12691059708595276, + 0.09352514892816544, + 0.05717974156141281, + 0.09523004293441772, + 0.03199518099427223, + 0.05805573984980583, + 0.2135394811630249, + 0.04670923575758934, + 0.07615236937999725 + ], + "credit_map": [ + 0.018002575263381004, + 0.022821001708507538, + 0.022816099226474762, + 0.03377728909254074, + 0.016845885664224625, + 0.02198832854628563, + 0.027505474165081978, + 0.03694585710763931, + 0.007458680309355259, + 0.012736347503960133, + 0.02744397334754467, + 0.008733833208680153, + 0.010333490557968616, + 0.01152864657342434, + 0.014722435735166073, + 0.03395318239927292, + 0.009474624879658222, + 0.010539498180150986, + 0.015795741230249405, + 0.01328131090849638, + 0.00921395979821682, + 0.00811243150383234, + 0.009303934872150421, + 0.017803648486733437, + 0.006698187906295061, + 0.008702033199369907, + 0.006848759017884731, + 0.006468231789767742, + 0.0070840767584741116, + 0.005951329134404659, + 0.010089129209518433, + 0.005337994545698166, + 0.007305794395506382, + 0.008183998055756092, + 0.012388214468955994, + 0.008573734201490879, + 0.013770261779427528, + 0.024212118238210678, + 0.013130808249115944, + 0.007665109355002642, + 0.0027618471067398787, + 0.0047531938180327415, + 0.005104124546051025, + 0.0034601856023073196, + 0.003561297431588173, + 0.0038440697826445103, + 0.00447771605104208, + 0.004032748285681009, + 0.00491336127743125, + 0.012698638252913952, + 0.006472982000559568, + 0.004870344419032335, + 0.009515667334198952, + 0.006342641543596983, + 0.008098602294921875, + 0.005143503192812204, + 0.02174542471766472, + 0.02267228253185749, + 0.028478719294071198, + 0.028725765645503998, + 0.02577262930572033, + 0.02613488957285881, + 0.025148889049887657, + 0.03486088290810585, + 0.005594729445874691, + 0.006689139641821384, + 0.0061447881162166595, + 0.00630352646112442, + 0.004827445838600397, + 0.00750760268419981, + 0.005372398532927036, + 0.004269605968147516, + 0.009581857360899448, + 0.007333726156502962, + 0.01513963658362627, + 0.006929914001375437, + 0.014001845382153988, + 0.008466261439025402, + 0.005308749154210091, + 0.00939038023352623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.020198959857225418 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.03373740240931511 + ], + "opposite_logprob_tokens": [ + -3.4087374210357666 + ], + "logprob_reward": -0.03373740240931511, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373740240931511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10276642441749573, + 0.24480195343494415, + 0.03278108313679695, + 0.06178586930036545, + 0.06775964796543121, + 0.059553779661655426, + 0.041454482823610306, + 0.2503933310508728, + 0.06318457424640656, + 0.07551883906126022 + ], + "credit_map": [ + 0.008819373324513435, + 0.01583029329776764, + 0.011381283402442932, + 0.020398147404193878, + 0.012058958411216736, + 0.011041003279387951, + 0.012264628894627094, + 0.010972734540700912, + 0.011003797873854637, + 0.017379827797412872, + 0.0645773708820343, + 0.016535190865397453, + 0.016476714983582497, + 0.02046944946050644, + 0.02168820984661579, + 0.07667139172554016, + 0.0029744235798716545, + 0.002962036756798625, + 0.005860405508428812, + 0.003958956804126501, + 0.004480977542698383, + 0.003034271067008376, + 0.004912348464131355, + 0.004597664810717106, + 0.00749977445229888, + 0.007697276771068573, + 0.010127454064786434, + 0.005240054335445166, + 0.007149812765419483, + 0.007408146280795336, + 0.01008873712271452, + 0.006574613507837057, + 0.006392086856067181, + 0.007476042956113815, + 0.007583377882838249, + 0.006891293916851282, + 0.010819918476045132, + 0.01179874874651432, + 0.00941434595733881, + 0.007383836433291435, + 0.006094395648688078, + 0.014711651019752026, + 0.008475342765450478, + 0.006906201131641865, + 0.005364508833736181, + 0.00599014712497592, + 0.00552792102098465, + 0.006483615376055241, + 0.0028339519631117582, + 0.005315402988344431, + 0.0034382252488285303, + 0.0038545080460608006, + 0.006543727591633797, + 0.008391872979700565, + 0.0072793494910001755, + 0.0037974445149302483, + 0.016550682485103607, + 0.01568579487502575, + 0.016346482560038567, + 0.015770679339766502, + 0.01629921793937683, + 0.07772766053676605, + 0.019775081425905228, + 0.07223773002624512, + 0.006989586632698774, + 0.007159960921853781, + 0.008257563225924969, + 0.009172767400741577, + 0.009405728429555893, + 0.00816658977419138, + 0.007580474019050598, + 0.006451900117099285, + 0.005774144548922777, + 0.009750093333423138, + 0.01200594287365675, + 0.008328965865075588, + 0.018241647630929947, + 0.00799620058387518, + 0.006246295291930437, + 0.00717554846778512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.03373740240931511 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.06209063529968262 + ], + "opposite_logprob_tokens": [ + -2.8120906352996826 + ], + "logprob_reward": -0.06209063529968262, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209063529968262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5203286409378052, + 0.03966962918639183, + 0.0716044008731842, + 0.02398568205535412, + 0.05452228710055351, + 0.03244956582784653, + 0.0639205127954483, + 0.09042579680681229, + 0.04229426011443138, + 0.0607992559671402 + ], + "credit_map": [ + 0.04158972203731537, + 0.05551042780280113, + 0.04867038503289223, + 0.08193068206310272, + 0.055678751319646835, + 0.06375361233949661, + 0.0898265540599823, + 0.08336851745843887, + 0.0030222118366509676, + 0.004747936502099037, + 0.007651088759303093, + 0.003034738125279546, + 0.004471112042665482, + 0.007279880810528994, + 0.0037943602073937654, + 0.005668299272656441, + 0.005821360275149345, + 0.008432267233729362, + 0.0063497996889054775, + 0.009236129932105541, + 0.01155609730631113, + 0.006491491105407476, + 0.01190431322902441, + 0.011812938377261162, + 0.002793343970552087, + 0.0024282224476337433, + 0.0027142614126205444, + 0.002294645644724369, + 0.002405875362455845, + 0.004777350928634405, + 0.003948487341403961, + 0.0026234956458210945, + 0.005164020229130983, + 0.006643315311521292, + 0.006593788508325815, + 0.004915302153676748, + 0.007534433621913195, + 0.011099981144070625, + 0.006963077466934919, + 0.005608365871012211, + 0.0028854524716734886, + 0.0062976679764688015, + 0.0038680920843034983, + 0.0035070900339633226, + 0.003279124153777957, + 0.0035317486617714167, + 0.005439571104943752, + 0.0036408184096217155, + 0.005586742423474789, + 0.00927257165312767, + 0.006443966180086136, + 0.009046094492077827, + 0.009282139129936695, + 0.007184737827628851, + 0.009912962093949318, + 0.007191296201199293, + 0.012978681363165379, + 0.009693792089819908, + 0.011196501553058624, + 0.011499698273837566, + 0.012078589759767056, + 0.010696056298911572, + 0.01108564157038927, + 0.011196834035217762, + 0.006234885659068823, + 0.008104780688881874, + 0.006104978732764721, + 0.004878230858594179, + 0.004323373548686504, + 0.0041642747819423676, + 0.004321336280554533, + 0.00416239770129323, + 0.0051499297842383385, + 0.006994093768298626, + 0.01101741287857294, + 0.006333237048238516, + 0.013996078632771969, + 0.007188169751316309, + 0.004430640954524279, + 0.005689688958227634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.06209063529968262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.16038963198661804 + ], + "opposite_logprob_tokens": [ + -1.9103896617889404 + ], + "logprob_reward": -0.16038963198661804, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038963198661804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060055654495954514, + 0.21739552915096283, + 0.051424961537122726, + 0.07259576767683029, + 0.05291646718978882, + 0.057378076016902924, + 0.03466860204935074, + 0.13853564858436584, + 0.12798310816287994, + 0.18704622983932495 + ], + "credit_map": [ + 0.005938160698860884, + 0.005069656763225794, + 0.006100247614085674, + 0.012917524203658104, + 0.0060817888006567955, + 0.008984079584479332, + 0.00745379738509655, + 0.007510395720601082, + 0.011545694433152676, + 0.02165079116821289, + 0.05202111601829529, + 0.014218715950846672, + 0.017788466066122055, + 0.01791614480316639, + 0.021311108022928238, + 0.06094348803162575, + 0.005639885552227497, + 0.006657099351286888, + 0.006714558694511652, + 0.005245323292911053, + 0.007629883009940386, + 0.004935149103403091, + 0.008929666131734848, + 0.005673395469784737, + 0.008020971901714802, + 0.012626060284674168, + 0.01259960513561964, + 0.008018812164664268, + 0.006761095020920038, + 0.00835192110389471, + 0.007865730673074722, + 0.008351565338671207, + 0.005648019723594189, + 0.009042125195264816, + 0.005849611945450306, + 0.007455231621861458, + 0.006543582770973444, + 0.0074811275117099285, + 0.006248723715543747, + 0.0046480377204716206, + 0.005692295264452696, + 0.012987198308110237, + 0.007393806241452694, + 0.008088301867246628, + 0.005264242645353079, + 0.0057604387402534485, + 0.006383483298122883, + 0.005808306857943535, + 0.004202721640467644, + 0.00479163508862257, + 0.004874261561781168, + 0.004088966641575098, + 0.005133340600878, + 0.004156768787652254, + 0.0040728249587118626, + 0.0033480790443718433, + 0.013073045760393143, + 0.015923406928777695, + 0.017190050333738327, + 0.014030959457159042, + 0.012193530797958374, + 0.023171033710241318, + 0.014240975491702557, + 0.02871263399720192, + 0.012298288755118847, + 0.021375156939029694, + 0.015776468440890312, + 0.015430464409291744, + 0.01427459716796875, + 0.012567784637212753, + 0.024421000853180885, + 0.011839335784316063, + 0.023589152842760086, + 0.019588612020015717, + 0.028202762827277184, + 0.019578801468014717, + 0.039214085787534714, + 0.018687674775719643, + 0.013911115936934948, + 0.024274008348584175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.16038963198661804 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.22557274997234344 + ], + "opposite_logprob_tokens": [ + -1.6005727052688599 + ], + "logprob_reward": -0.22557274997234344, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557274997234344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11713851988315582, + 0.07391899824142456, + 0.08207133412361145, + 0.05259270220994949, + 0.07416722923517227, + 0.03590516000986099, + 0.022136259824037552, + 0.33810892701148987, + 0.12349733710289001, + 0.08046354353427887 + ], + "credit_map": [ + 0.01483528595417738, + 0.010756973177194595, + 0.012428401969373226, + 0.014306738041341305, + 0.01103490311652422, + 0.018005944788455963, + 0.01599026285111904, + 0.019780011847615242, + 0.0036419727839529514, + 0.00718271778896451, + 0.014437032863497734, + 0.004730973858386278, + 0.008081790991127491, + 0.006352013908326626, + 0.011154115200042725, + 0.018338382244110107, + 0.00967023242264986, + 0.010952246375381947, + 0.011127743870019913, + 0.007531656417995691, + 0.010590209625661373, + 0.00915213581174612, + 0.013233190402388573, + 0.00981391966342926, + 0.0059646014124155045, + 0.01024513877928257, + 0.0066003380343317986, + 0.005727628245949745, + 0.005588514264672995, + 0.00449521467089653, + 0.006924103479832411, + 0.007047165650874376, + 0.008596060797572136, + 0.014751875773072243, + 0.008230352774262428, + 0.008168264292180538, + 0.006814876571297646, + 0.008837884292006493, + 0.010998677462339401, + 0.007769233547151089, + 0.0032521625980734825, + 0.006152365822345018, + 0.005318614188581705, + 0.0037502278573811054, + 0.0032821018248796463, + 0.005307279527187347, + 0.003694363171234727, + 0.005148044787347317, + 0.002328219823539257, + 0.0037323401775211096, + 0.0031541266944259405, + 0.0021432216744869947, + 0.0027445044834166765, + 0.0028599072247743607, + 0.0029893366154283285, + 0.002184602664783597, + 0.021443119272589684, + 0.021189482882618904, + 0.021867189556360245, + 0.02240113355219364, + 0.02247205562889576, + 0.10598445683717728, + 0.023727325722575188, + 0.09902416914701462, + 0.01597755402326584, + 0.01961076259613037, + 0.017181482166051865, + 0.015322654508054256, + 0.014424175955355167, + 0.013746336102485657, + 0.013526142574846745, + 0.013708222657442093, + 0.009079758077859879, + 0.009048582054674625, + 0.013665310107171535, + 0.01107096578925848, + 0.014369992539286613, + 0.010161909274756908, + 0.006622788961976767, + 0.006444231607019901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.22557274997234344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, handheld close-up captures the scene with visceral immediacy, the camera shaking dramatically in perfect sync with the action. A man's powerful hands dominate the frame; one grips the neck of an icy, brown glass beer bottle, smeared with condensation, while the other decisively hooks a simple stainless steel bottle opener beneath the metal cap. With a swift upward flick of the wrist, the cap pops free, and in that moment, the camera shakes, mimicking the force of the impact. Then, the man's hand casually drops the opener, its sound just out of focus, landing on the warm-toned wooden surface in the background. Without hesitation, he steadily pushes the bottle's mouth toward the camera, tilts it back, and takes a long, satisfying sip. The entire scene is bathed in the soft golden glow of the afternoon sun, casting long shadows and creating shimmers on the damp glass, evoking a relaxed, lighthearted atmosphere.", + "seed": 28410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_1/index_176_seed_28410_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": -0.07383861616253853, + "vlm_logprob": -0.07383861616253853, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.07383861616253853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.02020047791302204 + ], + "opposite_logprob_tokens": [ + -4.895200252532959 + ], + "logprob_reward": -0.02020047791302204, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.02020047791302204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2650918662548065, + 0.09466706216335297, + 0.07808931916952133, + 0.037376172840595245, + 0.12046832591295242, + 0.13854160904884338, + 0.039230331778526306, + 0.032369378954172134, + 0.14614547789096832, + 0.04802045226097107 + ], + "credit_map": [ + 0.02136417292058468, + 0.030122891068458557, + 0.04096205160021782, + 0.04032319784164429, + 0.028821107000112534, + 0.023801734670996666, + 0.05976128205657005, + 0.01993543654680252, + 0.010926639661192894, + 0.012458779849112034, + 0.012944784946739674, + 0.012308537028729916, + 0.01197627279907465, + 0.011637488380074501, + 0.01246615406125784, + 0.009948411025106907, + 0.00786871649324894, + 0.011012238450348377, + 0.008784298785030842, + 0.016846757382154465, + 0.006505361292511225, + 0.008194681257009506, + 0.011023030616343021, + 0.007854233495891094, + 0.00328054022975266, + 0.004029297269880772, + 0.004563014023005962, + 0.0048885648138821125, + 0.00506810937076807, + 0.004880531225353479, + 0.005785451736301184, + 0.004880667198449373, + 0.010203886777162552, + 0.01365438848733902, + 0.014933805912733078, + 0.018261900171637535, + 0.01487482525408268, + 0.018947722390294075, + 0.014569276943802834, + 0.015022513456642628, + 0.011269137263298035, + 0.01409065444022417, + 0.01374926045536995, + 0.010459274053573608, + 0.023889195173978806, + 0.03396566957235336, + 0.014503206126391888, + 0.016615213826298714, + 0.0038627092726528645, + 0.0028612890746444464, + 0.003878118237480521, + 0.005298570729792118, + 0.003991174977272749, + 0.008289354853332043, + 0.005958130583167076, + 0.005090986844152212, + 0.0035652199294418097, + 0.004369295667856932, + 0.0037186758127063513, + 0.004091208800673485, + 0.004802513401955366, + 0.005204277113080025, + 0.003475917736068368, + 0.0031422683969140053, + 0.01094383280724287, + 0.02154766023159027, + 0.01680549420416355, + 0.024901822209358215, + 0.020016146823763847, + 0.01781057007610798, + 0.020411496981978416, + 0.013708469457924366, + 0.007589682936668396, + 0.008073213510215282, + 0.005554480012506247, + 0.006175269838422537, + 0.007601125631481409, + 0.004232423380017281, + 0.004366154316812754, + 0.0044281031005084515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.02020047791302204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.023332679644227028 + ], + "opposite_logprob_tokens": [ + -3.7733325958251953 + ], + "logprob_reward": -0.023332679644227028, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023332679644227028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25114181637763977, + 0.16672702133655548, + 0.0367867648601532, + 0.13725309073925018, + 0.12704987823963165, + 0.06943430006504059, + 0.02754136361181736, + 0.06989482045173645, + 0.027532676234841347, + 0.08663829416036606 + ], + "credit_map": [ + 0.01917371340095997, + 0.025232430547475815, + 0.036434613168239594, + 0.0336913987994194, + 0.031848739832639694, + 0.021390873938798904, + 0.06028910353779793, + 0.02308095060288906, + 0.01988730952143669, + 0.027134237810969353, + 0.016035016626119614, + 0.020494040101766586, + 0.018096594139933586, + 0.023648014292120934, + 0.020346084609627724, + 0.0210857130587101, + 0.003200715407729149, + 0.0067911711521446705, + 0.0032072365283966064, + 0.005311279557645321, + 0.005256041884422302, + 0.004074582830071449, + 0.0041883778758347034, + 0.004757359623908997, + 0.013634392991662025, + 0.023116810247302055, + 0.019949574023485184, + 0.017207708209753036, + 0.021915597841143608, + 0.01091998815536499, + 0.018770789727568626, + 0.011738229542970657, + 0.009294657036662102, + 0.014937533996999264, + 0.01878315769135952, + 0.020792314782738686, + 0.014338667504489422, + 0.014578841626644135, + 0.022392909973859787, + 0.011931798420846462, + 0.005564471706748009, + 0.007233366370201111, + 0.006326055154204369, + 0.005514112766832113, + 0.015078257769346237, + 0.01572718657553196, + 0.007326873950660229, + 0.006663976702839136, + 0.002168502425774932, + 0.002465433906763792, + 0.0031237644143402576, + 0.004312323871999979, + 0.0026227228809148073, + 0.004435520153492689, + 0.004062953870743513, + 0.00435014208778739, + 0.0068495627492666245, + 0.010625972412526608, + 0.0077942353673279285, + 0.00712227588519454, + 0.014083663932979107, + 0.00819427240639925, + 0.005344325676560402, + 0.009880506433546543, + 0.002053930191323161, + 0.0036029587499797344, + 0.003408567514270544, + 0.005023937672376633, + 0.0030272388830780983, + 0.002768730977550149, + 0.003156597027555108, + 0.004490714520215988, + 0.005947501864284277, + 0.01921483315527439, + 0.003950375132262707, + 0.009568576700985432, + 0.005752165801823139, + 0.0046992129646241665, + 0.031115787103772163, + 0.00638984702527523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023332679644227028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.12716856598854065 + ], + "opposite_logprob_tokens": [ + -2.127168655395508 + ], + "logprob_reward": -0.12716856598854065, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12716856598854065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09524710476398468, + 0.1549806296825409, + 0.02886849455535412, + 0.10048811137676239, + 0.06785982847213745, + 0.30707111954689026, + 0.049506139010190964, + 0.10579477250576019, + 0.04970567673444748, + 0.04047815129160881 + ], + "credit_map": [ + 0.009062769822776318, + 0.010627890937030315, + 0.012396523728966713, + 0.012596851214766502, + 0.011164738796651363, + 0.007844675332307816, + 0.022152895107865334, + 0.009400761686265469, + 0.01761670969426632, + 0.017772598192095757, + 0.018208716064691544, + 0.01813528500497341, + 0.022461961954832077, + 0.021894751116633415, + 0.023804936558008194, + 0.0150856738910079, + 0.0032669496722519398, + 0.0036442182026803493, + 0.0031661740504205227, + 0.005330118350684643, + 0.0032075559720396996, + 0.002992181573063135, + 0.004401411395519972, + 0.002859884174540639, + 0.007627278100699186, + 0.014557752758264542, + 0.015839437022805214, + 0.014639833010733128, + 0.014542099088430405, + 0.007905411534011364, + 0.015069731511175632, + 0.010306566953659058, + 0.004683587234467268, + 0.0071768369525671005, + 0.013357636518776417, + 0.009631249122321606, + 0.006900325883179903, + 0.005497759208083153, + 0.012955322861671448, + 0.0076571060344576836, + 0.02414294146001339, + 0.031426701694726944, + 0.028233328834176064, + 0.020306572318077087, + 0.056170206516981125, + 0.0781470313668251, + 0.03730390965938568, + 0.03134043887257576, + 0.005304554011672735, + 0.0033746040426194668, + 0.003871687687933445, + 0.004239918198436499, + 0.005481728818267584, + 0.008801504969596863, + 0.009631174616515636, + 0.008800966665148735, + 0.009415700100362301, + 0.014637809246778488, + 0.009109559468925, + 0.010152087546885014, + 0.02260793186724186, + 0.015956107527017593, + 0.005938269197940826, + 0.017977304756641388, + 0.002552889985963702, + 0.004962195176631212, + 0.0034323676954954863, + 0.005736140999943018, + 0.013169785961508751, + 0.009390437044203281, + 0.005852452479302883, + 0.004609408788383007, + 0.0027285043615847826, + 0.005640905350446701, + 0.002758195623755455, + 0.005222150590270758, + 0.004056649748235941, + 0.0035142209380865097, + 0.012770803645253181, + 0.0037867212668061256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.12716856598854065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.1603829264640808 + ], + "opposite_logprob_tokens": [ + -1.9103829860687256 + ], + "logprob_reward": -0.1603829264640808, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603829264640808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13694709539413452, + 0.07264124602079391, + 0.033836256712675095, + 0.19990190863609314, + 0.07438568770885468, + 0.12208623439073563, + 0.06645715981721878, + 0.12419529259204865, + 0.11985988169908524, + 0.04968925192952156 + ], + "credit_map": [ + 0.01056408416479826, + 0.013710994273424149, + 0.01930942013859749, + 0.01886923797428608, + 0.018136557191610336, + 0.011909459717571735, + 0.03157461807131767, + 0.012872726656496525, + 0.006822745315730572, + 0.00922152865678072, + 0.008882696740329266, + 0.009719313122332096, + 0.008968041278421879, + 0.0071947891265153885, + 0.012270350009202957, + 0.00956177432090044, + 0.002541820053011179, + 0.004786923062056303, + 0.0027462029829621315, + 0.00397335272282362, + 0.005528098903596401, + 0.004566095303744078, + 0.004701727535575628, + 0.004992031957954168, + 0.017703628167510033, + 0.031133422628045082, + 0.03199668973684311, + 0.028549425303936005, + 0.030571842566132545, + 0.015240909531712532, + 0.024141773581504822, + 0.020564211532473564, + 0.006440776865929365, + 0.007750336546450853, + 0.012882838025689125, + 0.008123492822051048, + 0.011441787704825401, + 0.007666265591979027, + 0.012044323608279228, + 0.008035862818360329, + 0.009719348512589931, + 0.011852954514324665, + 0.012907206080853939, + 0.017711812630295753, + 0.015985408797860146, + 0.024692587554454803, + 0.01500344555824995, + 0.01421346329152584, + 0.006650872994214296, + 0.0057365261018276215, + 0.007063387893140316, + 0.007654125802218914, + 0.008917523548007011, + 0.0102246617898345, + 0.010607386939227581, + 0.009602673351764679, + 0.013384060934185982, + 0.014729839749634266, + 0.01121828705072403, + 0.018789591267704964, + 0.017771461978554726, + 0.022568073123693466, + 0.010934002697467804, + 0.014799967408180237, + 0.009765325114130974, + 0.01787097007036209, + 0.013225634582340717, + 0.02157128043472767, + 0.01517076138406992, + 0.018034053966403008, + 0.012372934259474277, + 0.01184891164302826, + 0.003886400954797864, + 0.012564782053232193, + 0.003832342801615596, + 0.005954445339739323, + 0.006872138008475304, + 0.00518537824973464, + 0.005534192081540823, + 0.005859570112079382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.1603829264640808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 tight handheld close-up featuring a man's hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an icy brown glass beer bottle with condensation visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man use a stainless steel bottle opener to pop the cap off the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shake dramatically in sync with the action of opening the bottle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft golden afternoon sunlight with long 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": [ + -0.03810843080282211 + ], + "opposite_logprob_tokens": [ + -3.2881083488464355 + ], + "logprob_reward": -0.03810843080282211, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810843080282211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0957363024353981, + 0.32175469398498535, + 0.06323426216840744, + 0.1466732770204544, + 0.0460871085524559, + 0.12223593890666962, + 0.03703987970948219, + 0.03169890493154526, + 0.05185573548078537, + 0.08368389308452606 + ], + "credit_map": [ + 0.008505270816385746, + 0.012045447714626789, + 0.0128811439499259, + 0.015577650628983974, + 0.011278746649622917, + 0.010495775379240513, + 0.016212575137615204, + 0.008739697746932507, + 0.03211401775479317, + 0.040996015071868896, + 0.033303581178188324, + 0.04244040325284004, + 0.03304601088166237, + 0.03788089379668236, + 0.054661061614751816, + 0.047312717884778976, + 0.007877703756093979, + 0.004805432632565498, + 0.007285911589860916, + 0.011953270062804222, + 0.00493483804166317, + 0.007717128377407789, + 0.008366623893380165, + 0.010293353348970413, + 0.012722447514533997, + 0.023489223793148994, + 0.023985134437680244, + 0.019769946113228798, + 0.02147485315799713, + 0.011564750224351883, + 0.02051381580531597, + 0.013153094798326492, + 0.005762667860835791, + 0.005517939105629921, + 0.006252505350857973, + 0.0059928190894424915, + 0.006631178315728903, + 0.004992527887225151, + 0.005082650110125542, + 0.005854817572981119, + 0.012300390750169754, + 0.011859724298119545, + 0.012622209265828133, + 0.013150660321116447, + 0.017976121976971626, + 0.03060690313577652, + 0.010903315618634224, + 0.012816622853279114, + 0.00477504450827837, + 0.0029764019418507814, + 0.0035603810101747513, + 0.004218022804707289, + 0.003986017312854528, + 0.005023727659136057, + 0.008116134442389011, + 0.004384150728583336, + 0.0040909890085458755, + 0.00476357014849782, + 0.002694906434044242, + 0.004361813422292471, + 0.0047788359224796295, + 0.004781000781804323, + 0.0033536318223923445, + 0.002874158788472414, + 0.0032247889321297407, + 0.005148919299244881, + 0.0033924703020602465, + 0.006135667208582163, + 0.009364227764308453, + 0.009596416726708412, + 0.006820764858275652, + 0.00817248597741127, + 0.004084158688783646, + 0.016456296667456627, + 0.003604314988479018, + 0.010917979292571545, + 0.006555251311510801, + 0.00502750976011157, + 0.031272899359464645, + 0.005765482783317566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.03810843080282211 + } + ], + "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_10/index_tensor([8621, 8621])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_tensor([8621, 8621])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e53e433338f1a7b816fbacc1e7667f45f2cd31ee --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_tensor([8621, 8621])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Animated in a whimsical Ghibli style, this eye-level medium shot gently tilts right to follow a red panda through a lush, hand-painted bamboo forest. The camera captures the creature's every joyful movement as it engages in a playful chase with falling cherry blossom petals. Soft, diffused natural light filters through the lush green canopy above, casting speckled sunspots on the forest floor and highlighting the rich, velvety texture of the red panda's reddish-brown fur. Its long, bushy tail, with a distinct cream-colored ring, sways with fluid grace. The background is a vibrant watercolor tapestry of layered greens and earthy browns with soft, blended edges. The creature, expressive and wide-eyed, gleams with pure joy, crouches down, then gently pounces, rolling onto a bed of soft fallen leaves and petals. Its front paws whimsically patter the swirling pink and white blossoms, creating an atmosphere of tranquility and enchantment.", + "seed": 30035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_8621_seed_30035_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": -0.01742947902530432, + "vlm_logprob": -0.01742947902530432, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.01742947902530432, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.007291254587471485 + ], + "opposite_logprob_tokens": [ + -5.507291316986084 + ], + "logprob_reward": -0.007291254587471485, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007291254587471485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12673957645893097, + 0.03230268135666847, + 0.030411453917622566, + 0.02704119123518467, + 0.43268176913261414, + 0.08337850868701935, + 0.11834612488746643, + 0.08382122218608856, + 0.04193844273686409, + 0.02333894930779934 + ], + "credit_map": [ + 0.011723325587809086, + 0.025172825902700424, + 0.016576427966356277, + 0.008318174630403519, + 0.011902800761163235, + 0.02870708331465721, + 0.0078017860651016235, + 0.016537174582481384, + 0.0056799957528710365, + 0.005965161137282848, + 0.004133689217269421, + 0.00281414482742548, + 0.003988479729741812, + 0.003166522365063429, + 0.0036454633809626102, + 0.0029092244803905487, + 0.0057270159013569355, + 0.005342578981071711, + 0.0038094292394816875, + 0.0027496779803186655, + 0.003311441745609045, + 0.003984589595347643, + 0.0028045857325196266, + 0.0026821368373930454, + 0.003914920147508383, + 0.006283894181251526, + 0.003791679861024022, + 0.0022547775879502296, + 0.0025969166308641434, + 0.0036480410490185022, + 0.0026494311168789864, + 0.0019015355501323938, + 0.14435090124607086, + 0.025132102891802788, + 0.1680973321199417, + 0.01644887588918209, + 0.011969143524765968, + 0.027999326586723328, + 0.012745724059641361, + 0.02593841403722763, + 0.008258258923888206, + 0.011623199097812176, + 0.008131100796163082, + 0.010766899213194847, + 0.013583759777247906, + 0.011876254342496395, + 0.012584091164171696, + 0.006554948631674051, + 0.013163868337869644, + 0.011431357823312283, + 0.02865464799106121, + 0.007133459206670523, + 0.01308287400752306, + 0.03120153211057186, + 0.00646212138235569, + 0.007216271944344044, + 0.011852610856294632, + 0.00974440760910511, + 0.011281108483672142, + 0.009907538071274757, + 0.012105213478207588, + 0.010192020796239376, + 0.00886499136686325, + 0.009873331524431705, + 0.006540108472108841, + 0.0043731569312512875, + 0.004202207084745169, + 0.004156129434704781, + 0.01134695578366518, + 0.0059945303946733475, + 0.002504129894077778, + 0.002821224508807063, + 0.0027120430022478104, + 0.004039010964334011, + 0.001890670508146286, + 0.0041765798814594746, + 0.004604519810527563, + 0.0021684025414288044, + 0.0021381184924393892, + 0.001609603874385357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.007291254587471485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.005261383950710297 + ], + "opposite_logprob_tokens": [ + -5.255261421203613 + ], + "logprob_reward": -0.005261383950710297, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261383950710297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14008153975009918, + 0.10474216938018799, + 0.05372879281640053, + 0.05027053877711296, + 0.18727274239063263, + 0.14337019622325897, + 0.06604477018117905, + 0.14846576750278473, + 0.04643486067652702, + 0.05958855524659157 + ], + "credit_map": [ + 0.017618536949157715, + 0.0261361263692379, + 0.0205888282507658, + 0.016238681972026825, + 0.017719395458698273, + 0.015222808346152306, + 0.008810670115053654, + 0.017746493220329285, + 0.010304801166057587, + 0.012643496505916119, + 0.013571101240813732, + 0.01256371010094881, + 0.014082401990890503, + 0.014767572283744812, + 0.01687939465045929, + 0.009929700754582882, + 0.009279771707952023, + 0.007268735207617283, + 0.010261569172143936, + 0.0056690433993935585, + 0.0053594377823174, + 0.005485417786985636, + 0.006400026381015778, + 0.004004789516329765, + 0.00769250700250268, + 0.008665730245411396, + 0.009403654374182224, + 0.003320172894746065, + 0.005564787890762091, + 0.007285744417458773, + 0.0031182633247226477, + 0.005219684913754463, + 0.06531426310539246, + 0.011828619986772537, + 0.06805407255887985, + 0.007805783301591873, + 0.006264299154281616, + 0.013293039053678513, + 0.005977251101285219, + 0.008735430426895618, + 0.012180582620203495, + 0.02329324558377266, + 0.013881416991353035, + 0.015904026105999947, + 0.029782049357891083, + 0.019317369908094406, + 0.018226664513349533, + 0.010784856975078583, + 0.006985102780163288, + 0.005183076485991478, + 0.007973751053214073, + 0.005269843153655529, + 0.010490434244275093, + 0.016308102756738663, + 0.005086933728307486, + 0.008747532032430172, + 0.01595929265022278, + 0.014405976049602032, + 0.0135341240093112, + 0.012447143904864788, + 0.02068963088095188, + 0.0373946949839592, + 0.015149865299463272, + 0.018885044381022453, + 0.006306798662990332, + 0.0060660988092422485, + 0.003741733031347394, + 0.004833977669477463, + 0.009155216626822948, + 0.0068396953865885735, + 0.006195012480020523, + 0.0032963326666504145, + 0.013393878005445004, + 0.007967058569192886, + 0.004599474836140871, + 0.014761685393750668, + 0.006545132491737604, + 0.0040443455800414085, + 0.0046833534725010395, + 0.0035936329513788223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.005261383950710297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.016056988388299942 + ], + "opposite_logprob_tokens": [ + -4.141057014465332 + ], + "logprob_reward": -0.016056988388299942, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016056988388299942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10128402709960938, + 0.08638279885053635, + 0.07486698031425476, + 0.03448902815580368, + 0.1733483225107193, + 0.06109615042805672, + 0.08311134576797485, + 0.21680516004562378, + 0.08504603058099747, + 0.08357009291648865 + ], + "credit_map": [ + 0.010966723784804344, + 0.013945993967354298, + 0.016318736597895622, + 0.012990790419280529, + 0.01654360257089138, + 0.014385546557605267, + 0.008197089657187462, + 0.007935541681945324, + 0.009388178586959839, + 0.014081855304539204, + 0.011770999059081078, + 0.009540772065520287, + 0.012459347024559975, + 0.007630467414855957, + 0.013467195443809032, + 0.008043983019888401, + 0.009097319096326828, + 0.009988012723624706, + 0.014219739474356174, + 0.010043688118457794, + 0.007897939532995224, + 0.005535198841243982, + 0.010886485688388348, + 0.007198598235845566, + 0.004388159140944481, + 0.006065324880182743, + 0.004426047671586275, + 0.003869769163429737, + 0.004181325901299715, + 0.005316962022334337, + 0.0022277366369962692, + 0.004013701807707548, + 0.05634292587637901, + 0.01274311263114214, + 0.05629146844148636, + 0.008984279818832874, + 0.00938326958566904, + 0.012944784015417099, + 0.006576817482709885, + 0.01008166465908289, + 0.008942650631070137, + 0.011372516863048077, + 0.006892051547765732, + 0.007036973722279072, + 0.007995156571269035, + 0.007100004702806473, + 0.007027965504676104, + 0.004728832747787237, + 0.0077995797619223595, + 0.006051146425306797, + 0.014708894304931164, + 0.005797883961349726, + 0.0120408209040761, + 0.023072440177202225, + 0.006580451503396034, + 0.00706013198941946, + 0.02024909481406212, + 0.018468424677848816, + 0.02055469900369644, + 0.020387444645166397, + 0.03426653519272804, + 0.05915994569659233, + 0.020074177533388138, + 0.023644840344786644, + 0.011462663300335407, + 0.007008560933172703, + 0.00797518715262413, + 0.007670057471841574, + 0.017770221456885338, + 0.014611294493079185, + 0.010803231038153172, + 0.007744820788502693, + 0.020234841853380203, + 0.009237680584192276, + 0.0071908533573150635, + 0.015547911636531353, + 0.008654768578708172, + 0.006615926511585712, + 0.009256025776267052, + 0.006832082290202379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.016056988388299942 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.003613373264670372 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613373264670372, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613373264670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10971320420503616, + 0.045936647802591324, + 0.03078121319413185, + 0.042340394109487534, + 0.5067729353904724, + 0.06972198188304901, + 0.05900325998663902, + 0.08333136141300201, + 0.02688836306333542, + 0.02551056444644928 + ], + "credit_map": [ + 0.010415422730147839, + 0.017489004880189896, + 0.01088994462043047, + 0.007529828231781721, + 0.014763138256967068, + 0.02036304399371147, + 0.009341439232230186, + 0.018921390175819397, + 0.005689534824341536, + 0.008521180599927902, + 0.006317462306469679, + 0.005572005640715361, + 0.004292110446840525, + 0.004274318926036358, + 0.006142409052699804, + 0.005127630662173033, + 0.00474562169983983, + 0.004485993646085262, + 0.004350566305220127, + 0.002842675196006894, + 0.0037068251986056566, + 0.004418987315148115, + 0.0028498924802988768, + 0.0033806522842496634, + 0.004469599109143019, + 0.0067787254229187965, + 0.009024951606988907, + 0.002952110953629017, + 0.004889384843409061, + 0.007538080681115389, + 0.0032737827859818935, + 0.0034137610346078873, + 0.1814902424812317, + 0.02803758718073368, + 0.19469599425792694, + 0.018443726003170013, + 0.010850210674107075, + 0.034461427479982376, + 0.014116333797574043, + 0.024677462875843048, + 0.006638763006776571, + 0.010820189490914345, + 0.007036129478365183, + 0.007118536625057459, + 0.013122913427650928, + 0.00997819285839796, + 0.00927739031612873, + 0.005729868542402983, + 0.009634401649236679, + 0.004225783981382847, + 0.008018317632377148, + 0.0033852311316877604, + 0.007796301506459713, + 0.009196680970489979, + 0.0037423125468194485, + 0.013004232197999954, + 0.010860432870686054, + 0.009678015485405922, + 0.010040332563221455, + 0.008834677748382092, + 0.012601863592863083, + 0.0151042016223073, + 0.007194492034614086, + 0.009017354808747768, + 0.004221807233989239, + 0.003653654595836997, + 0.0026286097709089518, + 0.0029322742484509945, + 0.0040204208344221115, + 0.004171471577137709, + 0.0031303155701607466, + 0.0021298096980899572, + 0.0031305064912885427, + 0.0055383644066751, + 0.002158593386411667, + 0.0038224603049457073, + 0.005638515576720238, + 0.0021209819242358208, + 0.001809845445677638, + 0.001291298889555037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.003613373264670372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.05492439493536949 + ], + "opposite_logprob_tokens": [ + -2.929924488067627 + ], + "logprob_reward": -0.05492439493536949, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492439493536949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06434071809053421, + 0.026964807882905006, + 0.02657523937523365, + 0.04181604087352753, + 0.44391486048698425, + 0.1504381000995636, + 0.0878371000289917, + 0.1069212555885315, + 0.044872455298900604, + 0.006319453939795494 + ], + "credit_map": [ + 0.0054258122108876705, + 0.012101303786039352, + 0.0077805230394005775, + 0.0034019658342003822, + 0.006251289043575525, + 0.014103399589657784, + 0.004983522929251194, + 0.010292897932231426, + 0.0031906254589557648, + 0.003681553527712822, + 0.003394490107893944, + 0.0029462198726832867, + 0.0032904164399951696, + 0.0025942607317119837, + 0.0046798293478786945, + 0.003187412628903985, + 0.004940744023770094, + 0.004200396127998829, + 0.005338476039469242, + 0.002688369480893016, + 0.002953600138425827, + 0.002591344527900219, + 0.002041986444965005, + 0.0018203210784122348, + 0.005114056635648012, + 0.009605631232261658, + 0.0070509156212210655, + 0.0029084612615406513, + 0.005367257632315159, + 0.005896621383726597, + 0.0031744816806167364, + 0.0026986184529960155, + 0.1456720530986786, + 0.028153253719210625, + 0.17203056812286377, + 0.01706749200820923, + 0.01111987791955471, + 0.02803223952651024, + 0.01327280793339014, + 0.028566569089889526, + 0.012406858615577221, + 0.025683801621198654, + 0.014269710518419743, + 0.016316542401909828, + 0.034323614090681076, + 0.019490379840135574, + 0.017877038568258286, + 0.010070158168673515, + 0.008497343398630619, + 0.0074486746452748775, + 0.01589510217308998, + 0.005474152509123087, + 0.011988937854766846, + 0.025369815528392792, + 0.004751707427203655, + 0.008411362767219543, + 0.014715509489178658, + 0.01345327403396368, + 0.013566840440034866, + 0.012802240438759327, + 0.015971088781952858, + 0.013611571863293648, + 0.010893560014665127, + 0.011907164007425308, + 0.006825941614806652, + 0.004992320202291012, + 0.004782773554325104, + 0.005059710703790188, + 0.008603199385106564, + 0.00716349296271801, + 0.0035886233672499657, + 0.003856392577290535, + 0.001224437030032277, + 0.000915226642973721, + 0.0005102711147628725, + 0.00079808768350631, + 0.0010700622806325555, + 0.0005290931439958513, + 0.0007406009244732559, + 0.0005316758179105818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.05492439493536949 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Animated in a whimsical Ghibli style, this eye-level medium shot gently tilts right to follow a red panda through a lush, hand-painted bamboo forest. The camera captures the creature's every joyful movement as it engages in a playful chase with falling cherry blossom petals. Soft, diffused natural light filters through the lush green canopy above, casting speckled sunspots on the forest floor and highlighting the rich, velvety texture of the red panda's reddish-brown fur. Its long, bushy tail, with a distinct cream-colored ring, sways with fluid grace. The background is a vibrant watercolor tapestry of layered greens and earthy browns with soft, blended edges. The creature, expressive and wide-eyed, gleams with pure joy, crouches down, then gently pounces, rolling onto a bed of soft fallen leaves and petals. Its front paws whimsically patter the swirling pink and white blossoms, creating an atmosphere of tranquility and enchantment.", + "seed": 30036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_8621_seed_30036_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": -0.018335899617522956, + "vlm_logprob": -0.018335899617522956, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.018335899617522956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.006843462120741606 + ], + "opposite_logprob_tokens": [ + -5.381843566894531 + ], + "logprob_reward": -0.006843462120741606, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006843462120741606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17464084923267365, + 0.05772863328456879, + 0.1555725783109665, + 0.06327854096889496, + 0.11364307254552841, + 0.07914260029792786, + 0.06704134494066238, + 0.09406150877475739, + 0.07328761368989944, + 0.12160325050354004 + ], + "credit_map": [ + 0.03992864117026329, + 0.022393671795725822, + 0.01695580966770649, + 0.019931919872760773, + 0.016217488795518875, + 0.02893398515880108, + 0.017751114442944527, + 0.01252823043614626, + 0.006510662380605936, + 0.006437020376324654, + 0.006791948806494474, + 0.00651862658560276, + 0.009162111207842827, + 0.007274233736097813, + 0.006827821489423513, + 0.00820621382445097, + 0.020514482632279396, + 0.017826814204454422, + 0.02010471001267433, + 0.015587922185659409, + 0.017834888771176338, + 0.025535831227898598, + 0.024175170809030533, + 0.01399277150630951, + 0.006858726963400841, + 0.007911995984613895, + 0.007879293523728848, + 0.010789701715111732, + 0.00867592915892601, + 0.005580910947173834, + 0.00907033309340477, + 0.006511657033115625, + 0.022044740617275238, + 0.01088614109903574, + 0.011366116814315319, + 0.012624325230717659, + 0.01586318016052246, + 0.019260205328464508, + 0.008098931983113289, + 0.013499438762664795, + 0.0091934222728014, + 0.009459522552788258, + 0.01299858558923006, + 0.012298417277634144, + 0.006515247281640768, + 0.00701116630807519, + 0.007353455759584904, + 0.01431279256939888, + 0.013500033877789974, + 0.004895469173789024, + 0.01299337763339281, + 0.008728931657969952, + 0.005553246941417456, + 0.005847962107509375, + 0.009551159106194973, + 0.005971166770905256, + 0.009496841579675674, + 0.007781162858009338, + 0.01113702543079853, + 0.014133515767753124, + 0.012997366487979889, + 0.015047463588416576, + 0.013731868006289005, + 0.009736272506415844, + 0.008098708465695381, + 0.0075957560911774635, + 0.011538737453520298, + 0.011378545314073563, + 0.007580242585390806, + 0.005443471949547529, + 0.012820149771869183, + 0.008832007646560669, + 0.011123730801045895, + 0.012826009653508663, + 0.009372318163514137, + 0.019196830689907074, + 0.019460024312138557, + 0.016964159905910492, + 0.01706230267882347, + 0.015597885474562645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.006843462120741606 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.011086177080869675 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011086177080869675, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086177080869675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13979007303714752, + 0.05070920288562775, + 0.20727910101413727, + 0.054983045905828476, + 0.10985887050628662, + 0.06072259321808815, + 0.05838736519217491, + 0.21344229578971863, + 0.038360338658094406, + 0.06646712869405746 + ], + "credit_map": [ + 0.009676069021224976, + 0.010691368952393532, + 0.012577665969729424, + 0.012948371469974518, + 0.05011199787259102, + 0.029655883088707924, + 0.009476185776293278, + 0.004652523901313543, + 0.006067393813282251, + 0.00528448261320591, + 0.005366909317672253, + 0.006313437595963478, + 0.00801130197942257, + 0.00650797551497817, + 0.007587425876408815, + 0.005570274777710438, + 0.02154470421373844, + 0.014997411519289017, + 0.025567706674337387, + 0.03971150144934654, + 0.035208720713853836, + 0.02652032859623432, + 0.030348192900419235, + 0.013380522839725018, + 0.005142244976013899, + 0.005593554116785526, + 0.006194653455168009, + 0.008444010280072689, + 0.008583692833781242, + 0.005594913382083178, + 0.010083835572004318, + 0.00534614035859704, + 0.009072905406355858, + 0.010405094362795353, + 0.010775547474622726, + 0.020274778828024864, + 0.014923003502190113, + 0.013751803897321224, + 0.01147459726780653, + 0.019181132316589355, + 0.004948372952640057, + 0.006868442986160517, + 0.012814194895327091, + 0.011846593581140041, + 0.005264937877655029, + 0.004066111985594034, + 0.0058602625504136086, + 0.009053681045770645, + 0.005597616080194712, + 0.012519062496721745, + 0.018519461154937744, + 0.005775536876171827, + 0.0042229886166751385, + 0.0029649091884493828, + 0.005154966376721859, + 0.0036328216083347797, + 0.01371291559189558, + 0.0088089220225811, + 0.06846991181373596, + 0.016705967485904694, + 0.011907578445971012, + 0.011186253279447556, + 0.01291593722999096, + 0.06973481923341751, + 0.003622812684625387, + 0.0032301442697644234, + 0.005187992472201586, + 0.00851008202880621, + 0.0033483090810477734, + 0.002850544173270464, + 0.0062509337440133095, + 0.005359518341720104, + 0.005146021954715252, + 0.011588753201067448, + 0.008410456590354443, + 0.01644607074558735, + 0.006267305463552475, + 0.005413603503257036, + 0.006527469493448734, + 0.006667449604719877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.011086177080869675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.014188130386173725 + ], + "opposite_logprob_tokens": [ + -4.264188289642334 + ], + "logprob_reward": -0.014188130386173725, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188130386173725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3443327248096466, + 0.04029051959514618, + 0.09535744041204453, + 0.07868877053260803, + 0.03918090835213661, + 0.09074461460113525, + 0.03951725363731384, + 0.17476022243499756, + 0.054506123065948486, + 0.042621467262506485 + ], + "credit_map": [ + 0.06009924039244652, + 0.021274736151099205, + 0.01176015380769968, + 0.025006553158164024, + 0.13541869819164276, + 0.057480450719594955, + 0.02436843141913414, + 0.00892441812902689, + 0.007015822920948267, + 0.00539960153400898, + 0.004568694159388542, + 0.0042663984932005405, + 0.004576277919113636, + 0.0044117518700659275, + 0.003231087001040578, + 0.0068208822049200535, + 0.016540104523301125, + 0.008155099116265774, + 0.01060379296541214, + 0.010873042978346348, + 0.008948818780481815, + 0.020175956189632416, + 0.011161205358803272, + 0.008899416774511337, + 0.00790160521864891, + 0.011070496402680874, + 0.006291399709880352, + 0.009909048676490784, + 0.011647660285234451, + 0.0077757020480930805, + 0.01660224050283432, + 0.007490607909858227, + 0.006496706511825323, + 0.0030349786393344402, + 0.0037827705964446068, + 0.004964924417436123, + 0.005767835304141045, + 0.0069369664415717125, + 0.0041538081131875515, + 0.004042913671582937, + 0.014799839816987514, + 0.012142840772867203, + 0.00788090843707323, + 0.011056832037866116, + 0.010456801392138004, + 0.010906270705163479, + 0.0077995131723582745, + 0.015701599419116974, + 0.003971203230321407, + 0.004168433137238026, + 0.004373590927571058, + 0.007720947265625, + 0.004544423893094063, + 0.0031316892709583044, + 0.007345182821154594, + 0.004261781461536884, + 0.01306634210050106, + 0.009378425776958466, + 0.04940563067793846, + 0.013555502519011497, + 0.014011302031576633, + 0.008106018416583538, + 0.013191303238272667, + 0.05404568463563919, + 0.0046210456639528275, + 0.004525479860603809, + 0.006983042694628239, + 0.011320844292640686, + 0.00440941471606493, + 0.00459646200761199, + 0.009382136166095734, + 0.008667693473398685, + 0.00437720213085413, + 0.0048632994294166565, + 0.004416408017277718, + 0.006191191263496876, + 0.008108639158308506, + 0.005914469715207815, + 0.004288628697395325, + 0.00446162698790431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.014188130386173725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.004635303281247616 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635303281247616, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635303281247616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11614952981472015, + 0.054752811789512634, + 0.11177589744329453, + 0.06604944914579391, + 0.10730624943971634, + 0.04135207086801529, + 0.034696221351623535, + 0.08889629691839218, + 0.07777625322341919, + 0.30124515295028687 + ], + "credit_map": [ + 0.010663294233381748, + 0.01254754327237606, + 0.009637015871703625, + 0.0197767186909914, + 0.024593710899353027, + 0.019980691373348236, + 0.012866300530731678, + 0.006084267981350422, + 0.007426073774695396, + 0.007478244602680206, + 0.006005808245390654, + 0.005455515347421169, + 0.010815908201038837, + 0.005484657362103462, + 0.004433706868439913, + 0.007652901578694582, + 0.01491352915763855, + 0.015104302205145359, + 0.011240758933126926, + 0.013929183594882488, + 0.01831706240773201, + 0.017259033396840096, + 0.013741455972194672, + 0.007270575501024723, + 0.007863545790314674, + 0.00928077194839716, + 0.007097352761775255, + 0.006888140458613634, + 0.006799772847443819, + 0.006472853012382984, + 0.012485803104937077, + 0.009161215275526047, + 0.016960402950644493, + 0.01160226110368967, + 0.008849007077515125, + 0.01590908318758011, + 0.01220121793448925, + 0.021745365113019943, + 0.007926471531391144, + 0.012112447060644627, + 0.0039026651065796614, + 0.007084541022777557, + 0.004525832831859589, + 0.005426988936960697, + 0.004427759442478418, + 0.0027785480488091707, + 0.0037029171362519264, + 0.009502824395895004, + 0.0030623245984315872, + 0.0038196765817701817, + 0.004556805361062288, + 0.005863984115421772, + 0.0036368584260344505, + 0.003024403238669038, + 0.00643040519207716, + 0.004301766399294138, + 0.00818480085581541, + 0.009284836240112782, + 0.008881837129592896, + 0.012792851775884628, + 0.014041407965123653, + 0.012525046244263649, + 0.013193405233323574, + 0.009992115199565887, + 0.007070981431752443, + 0.006995890289545059, + 0.007901644334197044, + 0.011456303298473358, + 0.01200321689248085, + 0.01461036130785942, + 0.007932319305837154, + 0.009805542416870594, + 0.02901405468583107, + 0.028747648000717163, + 0.022994212806224823, + 0.02834661491215229, + 0.05578510835766792, + 0.05883552506566048, + 0.031869977712631226, + 0.045652035623788834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.004635303281247616 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.05492642521858215 + ], + "opposite_logprob_tokens": [ + -2.9299263954162598 + ], + "logprob_reward": -0.05492642521858215, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492642521858215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39813122153282166, + 0.08877230435609818, + 0.07715083658695221, + 0.165117084980011, + 0.03051760420203209, + 0.046629227697849274, + 0.07306082546710968, + 0.05260618403553963, + 0.02399032935500145, + 0.044024329632520676 + ], + "credit_map": [ + 0.06565390527248383, + 0.026247741654515266, + 0.01309144590049982, + 0.0316244512796402, + 0.15507879853248596, + 0.06838276982307434, + 0.026909999549388885, + 0.011142143979668617, + 0.006701944395899773, + 0.017608970403671265, + 0.00606418726965785, + 0.008962627500295639, + 0.027094155550003052, + 0.007152676582336426, + 0.008965563960373402, + 0.00622218893840909, + 0.008820249699056149, + 0.00878726877272129, + 0.010301909409463406, + 0.013419123366475105, + 0.009500163607299328, + 0.010001488029956818, + 0.011431444436311722, + 0.004889199510216713, + 0.01866910047829151, + 0.023236572742462158, + 0.01782286912202835, + 0.022939113900065422, + 0.025179488584399223, + 0.016796628013253212, + 0.02636353112757206, + 0.014109793119132519, + 0.003787892870604992, + 0.003029622370377183, + 0.0035289875231683254, + 0.004763869568705559, + 0.004914205987006426, + 0.003986596595495939, + 0.002597239799797535, + 0.003909191116690636, + 0.006732507608830929, + 0.0051754070445895195, + 0.008733409456908703, + 0.006218982860445976, + 0.004769604187458754, + 0.005744648166000843, + 0.0032885717228055, + 0.005966097116470337, + 0.005311411805450916, + 0.008980807848274708, + 0.007360685151070356, + 0.017382793128490448, + 0.0073292008601129055, + 0.005794694181531668, + 0.012852142564952374, + 0.008049101568758488, + 0.005330324172973633, + 0.006499952636659145, + 0.00519159622490406, + 0.00698701199144125, + 0.010132345370948315, + 0.006739156786352396, + 0.006930537521839142, + 0.004795260727405548, + 0.0027064764872193336, + 0.0021384945139288902, + 0.0019594612531363964, + 0.0038616321980953217, + 0.004409002605825663, + 0.002781203482300043, + 0.0032544350251555443, + 0.002879624953493476, + 0.004023006651550531, + 0.008549380116164684, + 0.002903243526816368, + 0.0052702282555401325, + 0.00878836214542389, + 0.0064989496022462845, + 0.004048598930239677, + 0.0039425622671842575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.05492642521858215 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Animated in a whimsical Ghibli style, this eye-level medium shot gently tilts right to follow a red panda through a lush, hand-painted bamboo forest. The camera captures the creature's every joyful movement as it engages in a playful chase with falling cherry blossom petals. Soft, diffused natural light filters through the lush green canopy above, casting speckled sunspots on the forest floor and highlighting the rich, velvety texture of the red panda's reddish-brown fur. Its long, bushy tail, with a distinct cream-colored ring, sways with fluid grace. The background is a vibrant watercolor tapestry of layered greens and earthy browns with soft, blended edges. The creature, expressive and wide-eyed, gleams with pure joy, crouches down, then gently pounces, rolling onto a bed of soft fallen leaves and petals. Its front paws whimsically patter the swirling pink and white blossoms, creating an atmosphere of tranquility and enchantment.", + "seed": 30037, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_8621_seed_30037_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": -0.01848141774535179, + "vlm_logprob": -0.01848141774535179, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01848141774535179, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.012273500673472881 + ], + "opposite_logprob_tokens": [ + -4.51227331161499 + ], + "logprob_reward": -0.012273500673472881, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.012273500673472881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42320024967193604, + 0.10012969374656677, + 0.06667179614305496, + 0.0415322408080101, + 0.08429886400699615, + 0.045083846896886826, + 0.07366941124200821, + 0.033251844346523285, + 0.09465228021144867, + 0.03750968724489212 + ], + "credit_map": [ + 0.05238612741231918, + 0.04429600387811661, + 0.046923745423555374, + 0.04468417540192604, + 0.0472884364426136, + 0.042183201760053635, + 0.08855340629816055, + 0.05688519403338432, + 0.008128603920340538, + 0.008314166218042374, + 0.012574935331940651, + 0.010361405089497566, + 0.009429343976080418, + 0.01904449611902237, + 0.013765372335910797, + 0.018511390313506126, + 0.007311535999178886, + 0.007120003458112478, + 0.008102971129119396, + 0.006043469998985529, + 0.010478325188159943, + 0.007356132846325636, + 0.0077530439011752605, + 0.012506326660513878, + 0.004584236536175013, + 0.004605510737746954, + 0.005852042697370052, + 0.004707311280071735, + 0.007001454941928387, + 0.0043380712158977985, + 0.006109476555138826, + 0.004334143362939358, + 0.010309848934412003, + 0.01016892958432436, + 0.007403628434985876, + 0.014898085966706276, + 0.012626036070287228, + 0.008492453023791313, + 0.008532225154340267, + 0.011867668479681015, + 0.008042432367801666, + 0.004951133392751217, + 0.008351489901542664, + 0.0031671482138335705, + 0.004401963669806719, + 0.006264869123697281, + 0.0030435798689723015, + 0.006861233152449131, + 0.014238662086427212, + 0.007222251035273075, + 0.021254805848002434, + 0.005978506524115801, + 0.00572682311758399, + 0.005592557601630688, + 0.0048620314337313175, + 0.008793783374130726, + 0.005176384001970291, + 0.0041052186861634254, + 0.007390105631202459, + 0.002631127368658781, + 0.0027820803225040436, + 0.0033344931434839964, + 0.002687135012820363, + 0.0051453085616230965, + 0.013928616419434547, + 0.006738181225955486, + 0.01917092502117157, + 0.012846030294895172, + 0.010100885294377804, + 0.007838998921215534, + 0.008736074902117252, + 0.015292581170797348, + 0.00511664804071188, + 0.003486991161480546, + 0.0053651644848287106, + 0.004753488581627607, + 0.004517644643783569, + 0.0046622855588793755, + 0.005240035708993673, + 0.004367436282336712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.012273500673472881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.0032007675617933273 + ], + "opposite_logprob_tokens": [ + -5.753200531005859 + ], + "logprob_reward": -0.0032007675617933273, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032007675617933273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23691172897815704, + 0.07842504233121872, + 0.07720084488391876, + 0.06932998448610306, + 0.16213814914226532, + 0.0679120123386383, + 0.15144862234592438, + 0.042001690715551376, + 0.022190168499946594, + 0.09244172275066376 + ], + "credit_map": [ + 0.030626898631453514, + 0.03412781283259392, + 0.0314931683242321, + 0.041345685720443726, + 0.02226814441382885, + 0.017354540526866913, + 0.041972704231739044, + 0.01772278919816017, + 0.008060349151492119, + 0.007702123839408159, + 0.008795473724603653, + 0.004915766883641481, + 0.011876995675265789, + 0.015450780279934406, + 0.007834150455892086, + 0.013789398595690727, + 0.008037477731704712, + 0.007947610691189766, + 0.011750326491892338, + 0.012419549748301506, + 0.00804929155856371, + 0.009163783863186836, + 0.007451449520885944, + 0.012381356209516525, + 0.007819291204214096, + 0.009580169804394245, + 0.006408704910427332, + 0.006323613226413727, + 0.009745855815708637, + 0.008000178262591362, + 0.012609717436134815, + 0.008842448703944683, + 0.015081525780260563, + 0.02101317048072815, + 0.01700785756111145, + 0.02838391438126564, + 0.03207170218229294, + 0.011553115211427212, + 0.01821136102080345, + 0.018815504387021065, + 0.005774781107902527, + 0.00679277628660202, + 0.013883216306567192, + 0.008580651134252548, + 0.006545837037265301, + 0.0076248906552791595, + 0.008056220598518848, + 0.010653646662831306, + 0.02437238022685051, + 0.01541176252067089, + 0.04501013830304146, + 0.013175912201404572, + 0.012190558947622776, + 0.01202104240655899, + 0.0107532674446702, + 0.01851356215775013, + 0.005215827375650406, + 0.004520770162343979, + 0.00828061904758215, + 0.004366134759038687, + 0.004517102148383856, + 0.004701940808445215, + 0.0035491520538926125, + 0.006850146688520908, + 0.0027445799205452204, + 0.0018425901653245091, + 0.004191224928945303, + 0.0022203463595360518, + 0.002613351447507739, + 0.002661749254912138, + 0.0021941366139799356, + 0.0037221908569335938, + 0.01591951958835125, + 0.012366855517029762, + 0.013611857779324055, + 0.008484754711389542, + 0.009002831764519215, + 0.012457571923732758, + 0.010204371064901352, + 0.010393963195383549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0032007675617933273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.01606074348092079 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.01606074348092079, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606074348092079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2313285917043686, + 0.08148105442523956, + 0.05950013920664787, + 0.17846421897411346, + 0.06642420589923859, + 0.054063428193330765, + 0.029362333938479424, + 0.0409921258687973, + 0.04669444262981415, + 0.2116895318031311 + ], + "credit_map": [ + 0.030975330621004105, + 0.023033197969198227, + 0.027580292895436287, + 0.030979076400399208, + 0.02375408075749874, + 0.018323266878724098, + 0.051471270620822906, + 0.025212084874510765, + 0.008349884301424026, + 0.009814659133553505, + 0.012580681592226028, + 0.01013139821588993, + 0.007125095464289188, + 0.011485905386507511, + 0.008667203597724438, + 0.013326223008334637, + 0.007620806340128183, + 0.009841959923505783, + 0.010784308426082134, + 0.009997227229177952, + 0.004431437700986862, + 0.0065531241707503796, + 0.004983645863831043, + 0.005287628620862961, + 0.022161930799484253, + 0.023303980007767677, + 0.019101211801171303, + 0.01545667927712202, + 0.025609996169805527, + 0.020272448658943176, + 0.028311248868703842, + 0.02424672245979309, + 0.00809024553745985, + 0.00695978244766593, + 0.007767622824758291, + 0.010070673190057278, + 0.00772090395912528, + 0.006983804516494274, + 0.007062278222292662, + 0.011768895201385021, + 0.01047778595238924, + 0.003793518990278244, + 0.01611003838479519, + 0.004712591413408518, + 0.0036810042802244425, + 0.004963547457009554, + 0.003838570322841406, + 0.006486369762569666, + 0.006468450650572777, + 0.002824902068823576, + 0.005796099081635475, + 0.0028720973059535027, + 0.00224402011372149, + 0.0018949966179206967, + 0.0023406276013702154, + 0.004921140614897013, + 0.006096120923757553, + 0.0035360404290258884, + 0.009230701252818108, + 0.005155413877218962, + 0.004444832913577557, + 0.0036794249899685383, + 0.004312264267355204, + 0.004537329543381929, + 0.005423007998615503, + 0.004536857828497887, + 0.008334971964359283, + 0.005075725726783276, + 0.004233125131577253, + 0.004832736682146788, + 0.005355779081583023, + 0.008902236819267273, + 0.034467075020074844, + 0.026764128357172012, + 0.028620323166251183, + 0.018769411370158195, + 0.0233086496591568, + 0.026380492374300957, + 0.028083818033337593, + 0.02529563568532467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.01606074348092079 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.005946326069533825 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946326069533825, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946326069533825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19182045757770538, + 0.12994521856307983, + 0.18206839263439178, + 0.02653166651725769, + 0.09439054131507874, + 0.11453905701637268, + 0.03749540075659752, + 0.05386628955602646, + 0.1429288238286972, + 0.026414211839437485 + ], + "credit_map": [ + 0.025068487972021103, + 0.020971806719899178, + 0.024333395063877106, + 0.015326491557061672, + 0.023874010890722275, + 0.026451267302036285, + 0.03392389044165611, + 0.02187110297381878, + 0.011528999544680119, + 0.009965568780899048, + 0.013180178590118885, + 0.008919567801058292, + 0.022429432719945908, + 0.025852909311652184, + 0.01187829952687025, + 0.0261902566999197, + 0.016118448227643967, + 0.015097676776349545, + 0.030948009341955185, + 0.027137640863656998, + 0.018030676990747452, + 0.017896253615617752, + 0.018705451861023903, + 0.03813423216342926, + 0.004116714932024479, + 0.0037844327744096518, + 0.003345743054524064, + 0.0026984524447470903, + 0.002835251856595278, + 0.0031641186214983463, + 0.0034866409841924906, + 0.0031003111507743597, + 0.011277618817985058, + 0.013018983416259289, + 0.009429100900888443, + 0.01678646169602871, + 0.015373364090919495, + 0.009895743802189827, + 0.009780257008969784, + 0.008829006925225258, + 0.023441113531589508, + 0.009956653229892254, + 0.041472941637039185, + 0.00707005150616169, + 0.004975545220077038, + 0.008473288267850876, + 0.004686374217271805, + 0.014463089406490326, + 0.00697369547560811, + 0.004428577609360218, + 0.009504745714366436, + 0.0033706927206367254, + 0.002708361018449068, + 0.0030204863287508488, + 0.0023812686558812857, + 0.00510757602751255, + 0.008858900517225266, + 0.004222376272082329, + 0.019335061311721802, + 0.0042227753438055515, + 0.0038390844129025936, + 0.004291213117539883, + 0.004220517817884684, + 0.004876364022493362, + 0.022211462259292603, + 0.011021854355931282, + 0.028004413470625877, + 0.02062121592462063, + 0.016479436308145523, + 0.010661265812814236, + 0.012503235600888729, + 0.021425941959023476, + 0.004062142688781023, + 0.003683289047330618, + 0.005049428902566433, + 0.0023034042678773403, + 0.0023685263004153967, + 0.00295508885756135, + 0.002427759114652872, + 0.003564573125913739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.005946326069533825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.05492575094103813 + ], + "opposite_logprob_tokens": [ + -2.9299256801605225 + ], + "logprob_reward": -0.05492575094103813, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05492575094103813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1308162659406662, + 0.1383170634508133, + 0.362072229385376, + 0.08432412892580032, + 0.08786547183990479, + 0.05368490889668465, + 0.04913008585572243, + 0.04299933835864067, + 0.015294999815523624, + 0.03549555689096451 + ], + "credit_map": [ + 0.019573427736759186, + 0.01495463214814663, + 0.017772778868675232, + 0.014954677782952785, + 0.01585462875664234, + 0.01663866639137268, + 0.018985316157341003, + 0.012082135304808617, + 0.012283909134566784, + 0.011905994266271591, + 0.010792849585413933, + 0.009573116898536682, + 0.02339564450085163, + 0.027664029970765114, + 0.014551208354532719, + 0.02815030887722969, + 0.036379724740982056, + 0.035525158047676086, + 0.05686509236693382, + 0.05605195462703705, + 0.0695285052061081, + 0.04541026055812836, + 0.038338933140039444, + 0.02397257462143898, + 0.008786644786596298, + 0.010863754898309708, + 0.012647553347051144, + 0.009340951219201088, + 0.012489790096879005, + 0.008506703190505505, + 0.0107987429946661, + 0.010889980010688305, + 0.010417299345135689, + 0.011595455929636955, + 0.009014560841023922, + 0.014071113429963589, + 0.015212851576507092, + 0.010856952518224716, + 0.00726078636944294, + 0.009436452761292458, + 0.00579276867210865, + 0.004901583306491375, + 0.009138818830251694, + 0.006074942648410797, + 0.005412881262600422, + 0.009413649328052998, + 0.006061361636966467, + 0.006888903211802244, + 0.007787094451487064, + 0.005055703688412905, + 0.00882889237254858, + 0.004778583068400621, + 0.005630390252918005, + 0.007385249249637127, + 0.003933299798518419, + 0.00573087390512228, + 0.009702976793050766, + 0.00304813077673316, + 0.013350325636565685, + 0.0039822920225560665, + 0.002650040900334716, + 0.003374761203303933, + 0.0030394108034670353, + 0.003851397428661585, + 0.0020651633385568857, + 0.001599800307303667, + 0.0016917188186198473, + 0.0018717297352850437, + 0.0017636280972510576, + 0.002005685120820999, + 0.0015480488073080778, + 0.002749224891886115, + 0.005040381569415331, + 0.003653240855783224, + 0.005090991500765085, + 0.004504177253693342, + 0.004980356898158789, + 0.004029523581266403, + 0.00458893645554781, + 0.0036079497076570988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.05492575094103813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Animated in a whimsical Ghibli style, this eye-level medium shot gently tilts right to follow a red panda through a lush, hand-painted bamboo forest. The camera captures the creature's every joyful movement as it engages in a playful chase with falling cherry blossom petals. Soft, diffused natural light filters through the lush green canopy above, casting speckled sunspots on the forest floor and highlighting the rich, velvety texture of the red panda's reddish-brown fur. Its long, bushy tail, with a distinct cream-colored ring, sways with fluid grace. The background is a vibrant watercolor tapestry of layered greens and earthy browns with soft, blended edges. The creature, expressive and wide-eyed, gleams with pure joy, crouches down, then gently pounces, rolling onto a bed of soft fallen leaves and petals. Its front paws whimsically patter the swirling pink and white blossoms, creating an atmosphere of tranquility and enchantment.", + "seed": 30038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_10/index_8621_seed_30038_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": -0.02163660144433379, + "vlm_logprob": -0.02163660144433379, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.02163660144433379, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.008456612005829811 + ], + "opposite_logprob_tokens": [ + -5.133456707000732 + ], + "logprob_reward": -0.008456612005829811, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008456612005829811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18531271815299988, + 0.07613254338502884, + 0.0960083156824112, + 0.06485196202993393, + 0.06743110716342926, + 0.06913676857948303, + 0.257127046585083, + 0.07011111080646515, + 0.02734312228858471, + 0.08654526621103287 + ], + "credit_map": [ + 0.01913158781826496, + 0.035877302289009094, + 0.01753668300807476, + 0.013193430379033089, + 0.023206433281302452, + 0.02101079747080803, + 0.0236502718180418, + 0.03170620650053024, + 0.012750694528222084, + 0.007776552811264992, + 0.009694913402199745, + 0.006475203670561314, + 0.01554927695542574, + 0.009000157006084919, + 0.0070920116268098354, + 0.0077937268652021885, + 0.018594792112708092, + 0.010514480993151665, + 0.007247467525303364, + 0.008574191480875015, + 0.010185941122472286, + 0.008980466984212399, + 0.011928354389965534, + 0.019982626661658287, + 0.00728598702698946, + 0.009957028552889824, + 0.0068562086671590805, + 0.009682686999440193, + 0.008593101985752583, + 0.0086060781031847, + 0.0068207778967916965, + 0.007050097454339266, + 0.007134323474019766, + 0.006293395068496466, + 0.006978420540690422, + 0.00886556413024664, + 0.007954186759889126, + 0.008342370390892029, + 0.010673210956156254, + 0.011189636774361134, + 0.006877698004245758, + 0.01194430235773325, + 0.008305303752422333, + 0.005088169127702713, + 0.016084589064121246, + 0.0068793813697993755, + 0.0060179042629897594, + 0.007939416915178299, + 0.04593982547521591, + 0.04178965091705322, + 0.08766300231218338, + 0.01589001715183258, + 0.014597049914300442, + 0.019579259678721428, + 0.015432245098054409, + 0.016235990449786186, + 0.007984855212271214, + 0.009830959141254425, + 0.005733986850827932, + 0.008284398354589939, + 0.011917639523744583, + 0.011930747888982296, + 0.007462473586201668, + 0.006966050248593092, + 0.0055922833271324635, + 0.004651556722819805, + 0.0021513986866921186, + 0.002130575943738222, + 0.0027218724135309458, + 0.005506040062755346, + 0.00222949031740427, + 0.0023599062114953995, + 0.015229371376335621, + 0.01159188523888588, + 0.00774441147223115, + 0.008321280591189861, + 0.010217471048235893, + 0.013198679313063622, + 0.012234924361109734, + 0.008007247000932693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.008456612005829811 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.00195443257689476 + ], + "opposite_logprob_tokens": [ + -6.251954555511475 + ], + "logprob_reward": -0.00195443257689476, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00195443257689476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08673953264951706, + 0.3498845100402832, + 0.18421584367752075, + 0.03547047823667526, + 0.026315361261367798, + 0.08915343880653381, + 0.08393512666225433, + 0.03442123904824257, + 0.01747891679406166, + 0.09238560497760773 + ], + "credit_map": [ + 0.0063650282099843025, + 0.016527339816093445, + 0.007466763257980347, + 0.007234611548483372, + 0.014425323344767094, + 0.013280832208693027, + 0.009098057635128498, + 0.01234157383441925, + 0.08032317459583282, + 0.04335719347000122, + 0.0293316338211298, + 0.03265278413891792, + 0.063204325735569, + 0.03960047662258148, + 0.027841685339808464, + 0.0335732102394104, + 0.031409651041030884, + 0.02072870172560215, + 0.01881067082285881, + 0.021434905007481575, + 0.01918022334575653, + 0.013057726435363293, + 0.019704969599843025, + 0.039888981729745865, + 0.003217150457203388, + 0.00478658452630043, + 0.004775526467710733, + 0.0054209413938224316, + 0.005233221687376499, + 0.004882206209003925, + 0.0035610171034932137, + 0.0035938266664743423, + 0.0026346538215875626, + 0.0026421602815389633, + 0.002564274938777089, + 0.002614166820421815, + 0.0032069096341729164, + 0.003470666240900755, + 0.003477466059848666, + 0.00570506090298295, + 0.022114945575594902, + 0.00872084777802229, + 0.0065978397615253925, + 0.005678127985447645, + 0.025322770699858665, + 0.006746875122189522, + 0.007878742180764675, + 0.006093273870646954, + 0.00936911441385746, + 0.012838281691074371, + 0.010839670896530151, + 0.006832580082118511, + 0.008882085792720318, + 0.0072029875591397285, + 0.01583046466112137, + 0.012139934115111828, + 0.005960652604699135, + 0.005946771241724491, + 0.003134445520117879, + 0.003896466689184308, + 0.007756809704005718, + 0.002665685024112463, + 0.0028621654491871595, + 0.0021982414182275534, + 0.002059314167127013, + 0.0017194559331983328, + 0.0018733959877863526, + 0.002635106910020113, + 0.002597600221633911, + 0.001961920876055956, + 0.002268599346280098, + 0.002363521372899413, + 0.010695208795368671, + 0.01237720251083374, + 0.012353678233921528, + 0.006305258255451918, + 0.020466919988393784, + 0.011864985339343548, + 0.008512452244758606, + 0.009809893555939198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.00195443257689476 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.014188366010785103 + ], + "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": [ + 0.10305313020944595, + 0.08756238222122192, + 0.11565212160348892, + 0.036343928426504135, + 0.1809486448764801, + 0.0385291688144207, + 0.16603335738182068, + 0.03381701186299324, + 0.04757450520992279, + 0.19048573076725006 + ], + "credit_map": [ + 0.010364999994635582, + 0.021912548691034317, + 0.01061979215592146, + 0.011824582703411579, + 0.014614864252507687, + 0.010308021679520607, + 0.011051571927964687, + 0.01235676184296608, + 0.024550504982471466, + 0.011666557751595974, + 0.008618795312941074, + 0.006136971525847912, + 0.009666424244642258, + 0.01102622039616108, + 0.00729769142344594, + 0.008599217981100082, + 0.017675908282399178, + 0.013438134454190731, + 0.011807622388005257, + 0.017360690981149673, + 0.01197353471070528, + 0.010755425319075584, + 0.009962951764464378, + 0.022677863016724586, + 0.004026652313768864, + 0.004576715175062418, + 0.005866460036486387, + 0.006297192070633173, + 0.003848145017400384, + 0.005427887197583914, + 0.0031672853510826826, + 0.0031335914973169565, + 0.020787296816706657, + 0.024412672966718674, + 0.016431022435426712, + 0.018379854038357735, + 0.021386561915278435, + 0.024422382935881615, + 0.019292404875159264, + 0.03583645820617676, + 0.0053205182775855064, + 0.006454590708017349, + 0.003125485498458147, + 0.004538551904261112, + 0.005177798680961132, + 0.00407801428809762, + 0.00359605485573411, + 0.0062381550669670105, + 0.015576004981994629, + 0.041375864297151566, + 0.055044934153556824, + 0.010458116419613361, + 0.011768416501581669, + 0.011437331326305866, + 0.011404075659811497, + 0.008968625217676163, + 0.005818714387714863, + 0.005358378402888775, + 0.002759750699624419, + 0.0029657615814357996, + 0.007262163329869509, + 0.00331533863209188, + 0.0034407328348606825, + 0.002896170597523451, + 0.005574253853410482, + 0.0054794601164758205, + 0.00468461774289608, + 0.007286358624696732, + 0.008026576600968838, + 0.00548346946015954, + 0.006096402648836374, + 0.004943365231156349, + 0.02205766923725605, + 0.03938865289092064, + 0.022966964170336723, + 0.017712419852614403, + 0.023750610649585724, + 0.01989397406578064, + 0.0258595272898674, + 0.018855908885598183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.014188366010785103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.004637557547539473 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637557547539473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637557547539473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14217022061347961, + 0.208391934633255, + 0.11499578505754471, + 0.05357438698410988, + 0.037427544593811035, + 0.07367470860481262, + 0.16678819060325623, + 0.04467252269387245, + 0.02556568756699562, + 0.13273903727531433 + ], + "credit_map": [ + 0.012214316055178642, + 0.01701757311820984, + 0.013953350484371185, + 0.012725298292934895, + 0.019434958696365356, + 0.020581677556037903, + 0.020709792152047157, + 0.025533244013786316, + 0.049779798835515976, + 0.02590857818722725, + 0.017573917284607887, + 0.018951531499624252, + 0.03523516654968262, + 0.023557979613542557, + 0.01616424322128296, + 0.021220741793513298, + 0.025368988513946533, + 0.011112182401120663, + 0.009155196137726307, + 0.012760693207383156, + 0.008611232973635197, + 0.009235777892172337, + 0.00963591504842043, + 0.029115799814462662, + 0.005449960939586163, + 0.007857901975512505, + 0.00661774817854166, + 0.009342036210000515, + 0.0072694458067417145, + 0.006128568667918444, + 0.005771545693278313, + 0.005137179046869278, + 0.0053592766635119915, + 0.0038679137360304594, + 0.003589620813727379, + 0.004560929257422686, + 0.006675763987004757, + 0.003217861521989107, + 0.006419287994503975, + 0.0037368894554674625, + 0.020122965797781944, + 0.016815168783068657, + 0.006281764712184668, + 0.007563989143818617, + 0.01004430279135704, + 0.005506156478077173, + 0.0032756393775343895, + 0.004064719658344984, + 0.030225535854697227, + 0.03533695265650749, + 0.06859372556209564, + 0.008814213797450066, + 0.007909405045211315, + 0.006806965917348862, + 0.004992942791432142, + 0.004108448978513479, + 0.00825063232332468, + 0.008932647295296192, + 0.0033945178147405386, + 0.00621391087770462, + 0.007725112605839968, + 0.0036326111294329166, + 0.004196545574814081, + 0.002326544839888811, + 0.005307453218847513, + 0.0033979068975895643, + 0.0017951290355995297, + 0.0021381177939474583, + 0.0029465239495038986, + 0.0050376467406749725, + 0.002527007833123207, + 0.002415901981294155, + 0.016034001484513283, + 0.015854403376579285, + 0.013690554536879063, + 0.011854979209601879, + 0.02565387636423111, + 0.01762152835726738, + 0.019875837489962578, + 0.01215385366231203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.004637557547539473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a whimsical Ghibli style with a hand-painted watercolor aesthetic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red panda with reddish-brown fur and a long bushy tail featuring a cream-colored ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the red panda perform a sequence of crouching, pouncing, and rolling onto a bed of fallen leaves and petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a lush bamboo forest with falling cherry blossom petals and soft, diffused natural light creating speckled sunspots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start as an eye-level medium shot and gently tilt right to follow the red panda'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": [ + -0.07894603908061981 + ], + "opposite_logprob_tokens": [ + -2.578946113586426 + ], + "logprob_reward": -0.07894603908061981, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894603908061981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16366612911224365, + 0.34836041927337646, + 0.08811366558074951, + 0.055818285793066025, + 0.07232233136892319, + 0.02965044043958187, + 0.1233624592423439, + 0.0388806015253067, + 0.04170116409659386, + 0.038124457001686096 + ], + "credit_map": [ + 0.01418298576027155, + 0.019244572147727013, + 0.010877746157348156, + 0.008993086405098438, + 0.04099319130182266, + 0.029732266440987587, + 0.015075335279107094, + 0.02456696145236492, + 0.0733012706041336, + 0.04401491954922676, + 0.03185667842626572, + 0.03212321177124977, + 0.06650669127702713, + 0.0380430668592453, + 0.02984212152659893, + 0.03267247974872589, + 0.009607171639800072, + 0.00838436745107174, + 0.00857597030699253, + 0.014582150615751743, + 0.010281582362949848, + 0.010281099937856197, + 0.01208125427365303, + 0.01432006899267435, + 0.0068510170094668865, + 0.007723378017544746, + 0.005714872851967812, + 0.008558321744203568, + 0.009162798523902893, + 0.00736223952844739, + 0.004992503207176924, + 0.005453154910355806, + 0.0090866107493639, + 0.010964767076075077, + 0.00890213344246149, + 0.008980733342468739, + 0.009453412145376205, + 0.007597848307341337, + 0.009614100679755211, + 0.007722726557403803, + 0.004914639983326197, + 0.0030032964423298836, + 0.0037655956111848354, + 0.0024827320594340563, + 0.005010184366255999, + 0.0032368390820920467, + 0.0026627706829458475, + 0.004574383143335581, + 0.014406426809728146, + 0.03540710359811783, + 0.023310678079724312, + 0.008510347455739975, + 0.010566978715360165, + 0.01131220068782568, + 0.009418537840247154, + 0.010430189780890942, + 0.006540872622281313, + 0.006823447998613119, + 0.002454831264913082, + 0.0039116740226745605, + 0.0062583149410784245, + 0.0053088003769516945, + 0.004365962464362383, + 0.0032167003955692053, + 0.0065093934535980225, + 0.0037907799705863, + 0.0034037556033581495, + 0.004968069028109312, + 0.005803697742521763, + 0.008068099617958069, + 0.004758850671350956, + 0.004398517310619354, + 0.005345434881746769, + 0.008650721050798893, + 0.003269435605034232, + 0.0035692965611815453, + 0.004856667015701532, + 0.004028160125017166, + 0.0038659190759062767, + 0.004538823384791613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.07894603908061981 + } + ], + "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_11/index_tensor([7930, 7930])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_11/index_tensor([7930, 7930])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c4f61b59c87089707585c8934969511a1786b0f7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_11/index_tensor([7930, 7930])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid, artfully composed montage captures a teenage Pacific Islander girl twirling on a sunny beach, each moment a uniquely pictorial tableau. The sequence begins with a slow-motion medium shot, the camera gracefully arcing with her movements. She has long, wavy dark brown hair, a vibrant red hibiscus flower tucked behind her right ear. Her attire is a simple, lightweight sarong in a bright turquoise hue, its fabric billowing around her legs in the gentle sea breeze. As she twirls effortlessly on the damp, golden sand at the water's edge, the rapid cuts transition into a tight close-up of her face, a radiance of pure joy, her eyes crinkled in a genuine smile, her sun-kissed skin glowing warmly. The camera then tilts slightly downward as her bare feet kick up a fine mist of water and sand. The final shot in the series is a stunning wide shot, captured from a slightly lower angle, framing her mid-twirling, outstretched arms against the vast expanse of sparkling ocean and brilliant sunset with a breathtaking lens flare, bathing the entire scene in the warm, ethereal glow of the golden hour, evoking a powerful atmosphere of freedom, youth, and serene happiness.", + "seed": 19477, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_11/index_7930_seed_19477_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": -0.004282875312492251, + "vlm_logprob": -0.004282875312492251, + "vlm_yes_no_margin": 5.7, + "vlm_token_credit": -0.004282875312492251, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0068216766230762005 + ], + "opposite_logprob_tokens": [ + -5.756821632385254 + ], + "logprob_reward": -0.0068216766230762005, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0068216766230762005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1835804134607315, + 0.02967468649148941, + 0.5879144072532654, + 0.032518964260816574, + 0.03422842174768448, + 0.029568441212177277, + 0.01831730641424656, + 0.02857694774866104, + 0.028872353956103325, + 0.026747986674308777 + ], + "credit_map": [ + 0.06068130210042, + 0.010200271382927895, + 0.009443794377148151, + 0.007879513315856457, + 0.007454123813658953, + 0.016240069642663002, + 0.008413948118686676, + 0.06326739490032196, + 0.0027781836688518524, + 0.003912742715328932, + 0.004167409613728523, + 0.005705633200705051, + 0.003546631895005703, + 0.002012077486142516, + 0.004252584185451269, + 0.003299426520243287, + 0.24247467517852783, + 0.018711408600211143, + 0.021497288718819618, + 0.020721854642033577, + 0.012959020212292671, + 0.22756634652614594, + 0.013222156092524529, + 0.03076169453561306, + 0.0031993903685361147, + 0.00747382128611207, + 0.0047746640630066395, + 0.004341634456068277, + 0.003413146361708641, + 0.0030442539136856794, + 0.00264410930685699, + 0.0036279435735195875, + 0.004413587041199207, + 0.003974717576056719, + 0.005623714532703161, + 0.00394051056355238, + 0.00343657867051661, + 0.004082052037119865, + 0.003992857877165079, + 0.004764399025589228, + 0.0037215363699942827, + 0.0037782329600304365, + 0.0038139892276376486, + 0.002657931763678789, + 0.006089045200496912, + 0.0036658430472016335, + 0.0020238389261066914, + 0.0038180232513695955, + 0.0025587810669094324, + 0.0013494762824848294, + 0.0032480463851243258, + 0.002227790653705597, + 0.001289670355618, + 0.002837209263816476, + 0.00142296205740422, + 0.003383369417861104, + 0.0030122685711830854, + 0.0013255445519462228, + 0.0030233904253691435, + 0.0019736913964152336, + 0.0012060852022841573, + 0.007211270742118359, + 0.0027299211360514164, + 0.00809477549046278, + 0.002486689481884241, + 0.002618297701701522, + 0.005797623656690121, + 0.004005146212875843, + 0.0029183269944041967, + 0.004151810891926289, + 0.0027033600490540266, + 0.0041911001317203045, + 0.0021326583810150623, + 0.0037379744462668896, + 0.004932359326630831, + 0.005816933233290911, + 0.0016081216745078564, + 0.00312838819809258, + 0.0027858801186084747, + 0.002605672460049391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0068216766230762005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003627982921898365 + ], + "opposite_logprob_tokens": [ + -5.628627777099609 + ], + "logprob_reward": -0.003627982921898365, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003627982921898365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07433521747589111, + 0.02865745685994625, + 0.5307639837265015, + 0.03531968966126442, + 0.031624600291252136, + 0.06488237529993057, + 0.04818947613239288, + 0.02368493750691414, + 0.1508934646844864, + 0.011648782528936863 + ], + "credit_map": [ + 0.018351949751377106, + 0.009625021368265152, + 0.009041348472237587, + 0.00539952190592885, + 0.004131778609007597, + 0.0064132618717849255, + 0.0028766465838998556, + 0.01849568448960781, + 0.0030353846959769726, + 0.004113377537578344, + 0.004181866534054279, + 0.003987953066825867, + 0.0053100609220564365, + 0.00223100115545094, + 0.0033450175542384386, + 0.0024527935311198235, + 0.22223803400993347, + 0.01595000922679901, + 0.018107766285538673, + 0.017144933342933655, + 0.01111673191189766, + 0.2085193395614624, + 0.010075261816382408, + 0.027611924335360527, + 0.00371976918540895, + 0.006680793594568968, + 0.0037709525786340237, + 0.0045523876324296, + 0.0037361339200288057, + 0.004406115505844355, + 0.00285467435605824, + 0.00559886172413826, + 0.004101634956896305, + 0.006855628453195095, + 0.004449603147804737, + 0.004464833997189999, + 0.0029321082402020693, + 0.003481341991573572, + 0.0028850159142166376, + 0.0024544312618672848, + 0.01215387973934412, + 0.008272296749055386, + 0.013370941393077374, + 0.003813093528151512, + 0.004505494609475136, + 0.011053335852921009, + 0.003604382276535034, + 0.008108943700790405, + 0.005926452577114105, + 0.002917995909228921, + 0.007203380577266216, + 0.006367257796227932, + 0.0038160833064466715, + 0.007239297963678837, + 0.0049742236733436584, + 0.009744781069457531, + 0.003285334911197424, + 0.0017905941931530833, + 0.003559997770935297, + 0.00293944776058197, + 0.0016813284019008279, + 0.0032544711139053106, + 0.003819060744717717, + 0.0033547012135386467, + 0.008748212829232216, + 0.005902415607124567, + 0.04310160502791405, + 0.013834800571203232, + 0.006749794818460941, + 0.011437554843723774, + 0.008765083737671375, + 0.05235398933291435, + 0.0013361797900870442, + 0.0016120502259582281, + 0.0017037339275702834, + 0.0016418687300756574, + 0.0009678330970928073, + 0.0012563044438138604, + 0.001443956745788455, + 0.0016868533566594124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.003627982921898365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822227543219924 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822227543219924, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822227543219924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06794554740190506, + 0.03919374942779541, + 0.43626248836517334, + 0.027447331696748734, + 0.03096754290163517, + 0.029605068266391754, + 0.0444902665913105, + 0.07967718690633774, + 0.17091892659664154, + 0.07349186390638351 + ], + "credit_map": [ + 0.008801423013210297, + 0.0104265371337533, + 0.014449933543801308, + 0.0065299090929329395, + 0.005458689294755459, + 0.007072447333484888, + 0.006084210239350796, + 0.009122404269874096, + 0.004357659257948399, + 0.0057672723196446896, + 0.005575190298259258, + 0.006616658065468073, + 0.004730054177343845, + 0.004276243504136801, + 0.003990350756794214, + 0.0038803257048130035, + 0.18032486736774445, + 0.016666851937770844, + 0.01690421625971794, + 0.01306973397731781, + 0.009222396649420261, + 0.16919364035129547, + 0.008959830738604069, + 0.02192096598446369, + 0.003162729786708951, + 0.00403479253873229, + 0.00355124706402421, + 0.0038381305057555437, + 0.003190380986779928, + 0.00285527016967535, + 0.0026026226114481688, + 0.004212158266454935, + 0.0033676011953502893, + 0.01029190607368946, + 0.0022321774158626795, + 0.004154834430664778, + 0.003374004503712058, + 0.0030751130543649197, + 0.0022138769272714853, + 0.002258027670904994, + 0.0036470280028879642, + 0.004064794629812241, + 0.004603790584951639, + 0.0020027244463562965, + 0.0027496363036334515, + 0.004680328071117401, + 0.002429031766951084, + 0.0054277353920042515, + 0.005391399376094341, + 0.004804200958460569, + 0.0065085082314908504, + 0.004175134934484959, + 0.003640440758317709, + 0.007250669412314892, + 0.004993886686861515, + 0.007726026698946953, + 0.012976864352822304, + 0.004696335177868605, + 0.013975176028907299, + 0.00653082225471735, + 0.005501000676304102, + 0.011672409251332283, + 0.012666720896959305, + 0.011657855473458767, + 0.01530049741268158, + 0.011213352903723717, + 0.04207674413919449, + 0.015291858464479446, + 0.009188823401927948, + 0.01646711304783821, + 0.010052170604467392, + 0.051328375935554504, + 0.0073722661472857, + 0.01173825841397047, + 0.012084919959306717, + 0.01383285690099001, + 0.005273204762488604, + 0.0093931183218956, + 0.006069888360798359, + 0.0077273487113416195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.002822227543219924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.15032930672168732, + 0.0856412723660469, + 0.4481808543205261, + 0.05793525278568268, + 0.07005749642848969, + 0.03857653588056564, + 0.05347996577620506, + 0.029496071860194206, + 0.052142102271318436, + 0.014161127619445324 + ], + "credit_map": [ + 0.041647132486104965, + 0.013135850429534912, + 0.01812666654586792, + 0.007392411585897207, + 0.007693128660321236, + 0.014076098799705505, + 0.006344596855342388, + 0.04191342741250992, + 0.0075903162360191345, + 0.011023110710084438, + 0.010072773322463036, + 0.01257592998445034, + 0.013545007444918156, + 0.010249410755932331, + 0.00970538705587387, + 0.010879334062337875, + 0.1850137710571289, + 0.017429204657673836, + 0.015948446467518806, + 0.012988184578716755, + 0.009534303098917007, + 0.17295832931995392, + 0.010397480800747871, + 0.02391115203499794, + 0.004158945754170418, + 0.00859034713357687, + 0.004182273522019386, + 0.008200925774872303, + 0.011851785704493523, + 0.006479525472968817, + 0.00638243556022644, + 0.00808901060372591, + 0.008934801444411278, + 0.012549708597362041, + 0.008018571883440018, + 0.007991415448486805, + 0.0071492912247776985, + 0.00865664891898632, + 0.005970368627458811, + 0.010786696337163448, + 0.0045079803094267845, + 0.0037017061840742826, + 0.00648541422560811, + 0.0027396671939641237, + 0.003897366812452674, + 0.0063930111937224865, + 0.0030515077523887157, + 0.0077998824417591095, + 0.006572590675204992, + 0.0038125065620988607, + 0.006095620803534985, + 0.00615827227011323, + 0.0051280963234603405, + 0.007851563394069672, + 0.005402697250247002, + 0.01245861779898405, + 0.002957593183964491, + 0.0019311608048155904, + 0.004093509167432785, + 0.003981387242674828, + 0.0018609167309477925, + 0.0040291729383170605, + 0.005909173283725977, + 0.004733159206807613, + 0.00668995501473546, + 0.004690459929406643, + 0.009371662512421608, + 0.005213842261582613, + 0.002814811421558261, + 0.006781137082725763, + 0.004055127501487732, + 0.012525108642876148, + 0.0013280690182000399, + 0.0018152541015297174, + 0.0022580132354050875, + 0.002479337155818939, + 0.0008510233019478619, + 0.0017124820733442903, + 0.0012789351167157292, + 0.002438013441860676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.005946444813162088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196044661104679 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002196044661104679, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196044661104679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17054487764835358, + 0.029452873393893242, + 0.41132238507270813, + 0.02195669338107109, + 0.015609229914844036, + 0.017936114221811295, + 0.033200543373823166, + 0.11882719397544861, + 0.13743074238300323, + 0.043719399720430374 + ], + "credit_map": [ + 0.05377206578850746, + 0.009742611087858677, + 0.015419251285493374, + 0.006970727350562811, + 0.0053533283062279224, + 0.0166205782443285, + 0.0058077056892216206, + 0.0568586066365242, + 0.003753512864932418, + 0.0036741052754223347, + 0.0045371572487056255, + 0.003420908935368061, + 0.005515420343726873, + 0.0027153005357831717, + 0.00316045549698174, + 0.0026760133914649487, + 0.1717832237482071, + 0.012213998474180698, + 0.014706885442137718, + 0.01295139268040657, + 0.008691024966537952, + 0.16172648966312408, + 0.007750824559479952, + 0.02149854227900505, + 0.0022834271658211946, + 0.003672317834571004, + 0.0023329786490648985, + 0.0030028652399778366, + 0.002199151087552309, + 0.0027831848710775375, + 0.0019347023917362094, + 0.00374806672334671, + 0.0016142536187544465, + 0.0035499269142746925, + 0.0015604833606630564, + 0.0017419361975044012, + 0.002642624080181122, + 0.0014165309257805347, + 0.0015773013001307845, + 0.0015061723534017801, + 0.0030342470854520798, + 0.0020546705927699804, + 0.003067298559471965, + 0.0012611690908670425, + 0.001894684275612235, + 0.00274833501316607, + 0.0013544493122026324, + 0.0025212601758539677, + 0.005095647647976875, + 0.002361762337386608, + 0.00482017919421196, + 0.0036854471545666456, + 0.003360421396791935, + 0.004805597010999918, + 0.003055165521800518, + 0.0060163214802742004, + 0.008508133701980114, + 0.0038705633487552404, + 0.009658665396273136, + 0.005384428892284632, + 0.004184086807072163, + 0.034930769354104996, + 0.010951085016131401, + 0.0413394570350647, + 0.007903451099991798, + 0.003950548358261585, + 0.0423821322619915, + 0.01278498861938715, + 0.003261295612901449, + 0.011369479820132256, + 0.004245763644576073, + 0.051533084362745285, + 0.004495928529649973, + 0.007229941431432962, + 0.006071249023079872, + 0.007322987075895071, + 0.0040992069989442825, + 0.005717471241950989, + 0.0044570923782885075, + 0.004325525835156441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.002196044661104679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, artfully composed montage captures a teenage Pacific Islander girl twirling on a sunny beach, each moment a uniquely pictorial tableau. The sequence begins with a slow-motion medium shot, the camera gracefully arcing with her movements. She has long, wavy dark brown hair, a vibrant red hibiscus flower tucked behind her right ear. Her attire is a simple, lightweight sarong in a bright turquoise hue, its fabric billowing around her legs in the gentle sea breeze. As she twirls effortlessly on the damp, golden sand at the water's edge, the rapid cuts transition into a tight close-up of her face, a radiance of pure joy, her eyes crinkled in a genuine smile, her sun-kissed skin glowing warmly. The camera then tilts slightly downward as her bare feet kick up a fine mist of water and sand. The final shot in the series is a stunning wide shot, captured from a slightly lower angle, framing her mid-twirling, outstretched arms against the vast expanse of sparkling ocean and brilliant sunset with a breathtaking lens flare, bathing the entire scene in the warm, ethereal glow of the golden hour, evoking a powerful atmosphere of freedom, youth, and serene happiness.", + "seed": 19478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_11/index_7930_seed_19478_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": -0.047395438561215995, + "vlm_logprob": -0.047395438561215995, + "vlm_yes_no_margin": 5.325, + "vlm_token_credit": -0.047395438561215995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00533039728179574 + ], + "opposite_logprob_tokens": [ + -6.380330562591553 + ], + "logprob_reward": -0.00533039728179574, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.00533039728179574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052560243755578995, + 0.06732603162527084, + 0.37557360529899597, + 0.07137633860111237, + 0.13436655700206757, + 0.028889888897538185, + 0.059333495795726776, + 0.0659162700176239, + 0.0684320479631424, + 0.07622557133436203 + ], + "credit_map": [ + 0.0068481555208563805, + 0.004078757017850876, + 0.007521443068981171, + 0.005438561551272869, + 0.005147937219589949, + 0.008027360774576664, + 0.006601606495678425, + 0.008896423503756523, + 0.008234009146690369, + 0.007824834436178207, + 0.005867629311978817, + 0.01164704654365778, + 0.006060847081243992, + 0.009624681435525417, + 0.009041695855557919, + 0.009025290608406067, + 0.1244470626115799, + 0.01363820768892765, + 0.012803660705685616, + 0.00834848266094923, + 0.0175386443734169, + 0.02634577453136444, + 0.02382298745214939, + 0.14862878620624542, + 0.014517945237457752, + 0.0050192163325846195, + 0.010092792101204395, + 0.006061954889446497, + 0.004377106670290232, + 0.010503520257771015, + 0.00529584800824523, + 0.01550795417279005, + 0.01622634194791317, + 0.010641191154718399, + 0.020782217383384705, + 0.01818929985165596, + 0.02174357697367668, + 0.00881948508322239, + 0.022920088842511177, + 0.015044363215565681, + 0.0042678541503846645, + 0.0028506640810519457, + 0.004801598377525806, + 0.004828858654946089, + 0.0030297173652797937, + 0.0044520082883536816, + 0.0027023477014154196, + 0.0019568384159356356, + 0.007931342348456383, + 0.009952097199857235, + 0.007335517089813948, + 0.006848298478871584, + 0.009144448675215244, + 0.004323960747569799, + 0.006933237891644239, + 0.006864587310701609, + 0.010667284019291401, + 0.00738725857809186, + 0.013505501672625542, + 0.005909069441258907, + 0.008737869560718536, + 0.004362183157354593, + 0.008794556371867657, + 0.006552548613399267, + 0.01044234074652195, + 0.0036072093062102795, + 0.005708531476557255, + 0.004522842820733786, + 0.008603669703006744, + 0.02108602598309517, + 0.006146303843706846, + 0.008315120823681355, + 0.008742069825530052, + 0.011188618838787079, + 0.009953588247299194, + 0.008611710742115974, + 0.007158211898058653, + 0.008700498379766941, + 0.013153408654034138, + 0.008717458695173264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.00533039728179574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011863820254802704 + ], + "opposite_logprob_tokens": [ + -6.751186370849609 + ], + "logprob_reward": -0.0011863820254802704, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011863820254802704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039217736572027206, + 0.06930351257324219, + 0.4514921307563782, + 0.044199373573064804, + 0.04610629752278328, + 0.030923455953598022, + 0.10450597107410431, + 0.05744321271777153, + 0.1139078363776207, + 0.04290048032999039 + ], + "credit_map": [ + 0.004437096416950226, + 0.003542782738804817, + 0.00644204206764698, + 0.005092962179332972, + 0.003492791438475251, + 0.004816807806491852, + 0.004584598820656538, + 0.006808652076870203, + 0.007774148602038622, + 0.008573178201913834, + 0.011547238565981388, + 0.008590909652411938, + 0.007607598789036274, + 0.010238861665129662, + 0.006306284572929144, + 0.008665289729833603, + 0.1469595730304718, + 0.015872381627559662, + 0.01802392490208149, + 0.008400727063417435, + 0.021371014416217804, + 0.034669335931539536, + 0.027918213978409767, + 0.178276926279068, + 0.008435693569481373, + 0.0036698977928608656, + 0.006499988026916981, + 0.004320999141782522, + 0.0027110318187624216, + 0.00602302048355341, + 0.003907643724232912, + 0.008631094358861446, + 0.00390846561640501, + 0.00647450378164649, + 0.008775443769991398, + 0.009029073640704155, + 0.003900621086359024, + 0.002685276325792074, + 0.005420554429292679, + 0.005912356078624725, + 0.00512401107698679, + 0.003783242078498006, + 0.004080659709870815, + 0.0034473410341888666, + 0.0037071104161441326, + 0.004508089739829302, + 0.003842841600999236, + 0.0024301607627421618, + 0.013630921952426434, + 0.014467108994722366, + 0.011381620541214943, + 0.01206377986818552, + 0.015545197762548923, + 0.011556999757885933, + 0.012587090022861958, + 0.013273260556161404, + 0.007253422401845455, + 0.00974076334387064, + 0.008933808654546738, + 0.004651682451367378, + 0.00970013439655304, + 0.004605304915457964, + 0.007328880485147238, + 0.005229216534644365, + 0.01973986066877842, + 0.015166785567998886, + 0.009456002153456211, + 0.010813333094120026, + 0.008876163512468338, + 0.02746507339179516, + 0.00861179456114769, + 0.013778822496533394, + 0.0050314199179410934, + 0.0050843264907598495, + 0.004213757812976837, + 0.005224672611802816, + 0.007400644943118095, + 0.004966386128216982, + 0.006053224205970764, + 0.004926048684865236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.0011863820254802704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001714071026071906 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.001714071026071906, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001714071026071906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06600729376077652, + 0.06105782091617584, + 0.6256019473075867, + 0.03341525048017502, + 0.0598088800907135, + 0.01936137117445469, + 0.022135097533464432, + 0.042598456144332886, + 0.038781002163887024, + 0.031232867389917374 + ], + "credit_map": [ + 0.00790418777614832, + 0.00704540777951479, + 0.008433478884398937, + 0.006787007208913565, + 0.010095636360347271, + 0.009753835387527943, + 0.00891325157135725, + 0.007074498571455479, + 0.006780820433050394, + 0.004147747531533241, + 0.01401031855493784, + 0.005304093472659588, + 0.007658663671463728, + 0.01022262405604124, + 0.00544569781050086, + 0.007487863767892122, + 0.2051318883895874, + 0.022245850414037704, + 0.023168208077549934, + 0.011730820871889591, + 0.029649030417203903, + 0.04625564068555832, + 0.037758223712444305, + 0.2496623396873474, + 0.006737044546753168, + 0.0021447932813316584, + 0.004404784645885229, + 0.002233159961178899, + 0.003094422398135066, + 0.005037268623709679, + 0.0023278791923075914, + 0.007435903884470463, + 0.00789925828576088, + 0.005421728827059269, + 0.01047517266124487, + 0.006271279416978359, + 0.00769141037017107, + 0.005188118666410446, + 0.007561507169157267, + 0.009300412610173225, + 0.0024411447811871767, + 0.0023902812972664833, + 0.0016868581296876073, + 0.002090912079438567, + 0.0028860019519925117, + 0.0028966639656573534, + 0.003070034785196185, + 0.001899476395919919, + 0.0029216536786407232, + 0.002446262165904045, + 0.0032726996578276157, + 0.0021585565991699696, + 0.00437859958037734, + 0.0021582043264061213, + 0.002775894943624735, + 0.0020232289098203182, + 0.007225772365927696, + 0.006010416429489851, + 0.007697097025811672, + 0.00380098563618958, + 0.005944297648966312, + 0.003029389539733529, + 0.005874376278370619, + 0.0030161242466419935, + 0.0072837271727621555, + 0.004163255449384451, + 0.004906800109893084, + 0.003653095569461584, + 0.0035577553790062666, + 0.0049789417535066605, + 0.0037627127021551132, + 0.00647471658885479, + 0.0033564900513738394, + 0.004386832471936941, + 0.003287149593234062, + 0.0036664640065282583, + 0.004945883993059397, + 0.0033177039586007595, + 0.004420188255608082, + 0.0038521592505276203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.001714071026071906 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189953975379467 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003189953975379467, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189953975379467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07058928906917572, + 0.07926081866025925, + 0.0447230227291584, + 0.04996639862656593, + 0.2111119031906128, + 0.06238631159067154, + 0.053963638842105865, + 0.06259818375110626, + 0.11746223270893097, + 0.2479381263256073 + ], + "credit_map": [ + 0.008839874528348446, + 0.0052266232669353485, + 0.013795568607747555, + 0.006385122425854206, + 0.007078900467604399, + 0.009382792748510838, + 0.00781219732016325, + 0.012068213894963264, + 0.007266090251505375, + 0.009279722347855568, + 0.015288243070244789, + 0.008900425396859646, + 0.011197065003216267, + 0.011088546365499496, + 0.0064834114164114, + 0.009757314808666706, + 0.00629673944786191, + 0.005967685952782631, + 0.00498546939343214, + 0.006336181424558163, + 0.004495497327297926, + 0.004918212071061134, + 0.003943734336644411, + 0.007779504172503948, + 0.009103764779865742, + 0.0036375103518366814, + 0.006156358402222395, + 0.005241393577307463, + 0.005035589449107647, + 0.006433312781155109, + 0.004654808435589075, + 0.009703664109110832, + 0.019087443128228188, + 0.019716350361704826, + 0.0333152636885643, + 0.04723405838012695, + 0.013531182892620564, + 0.012181058526039124, + 0.028033025562763214, + 0.0380135215818882, + 0.00902682263404131, + 0.006085275672376156, + 0.008868023753166199, + 0.00840994156897068, + 0.006063255947083235, + 0.009399570524692535, + 0.010084589943289757, + 0.004448828287422657, + 0.006734086666256189, + 0.005786170717328787, + 0.007650029379874468, + 0.006713288836181164, + 0.005780296865850687, + 0.010171957314014435, + 0.004693974740803242, + 0.006433829665184021, + 0.009416320361196995, + 0.009004319086670876, + 0.009647553786635399, + 0.006695229560136795, + 0.009377661161124706, + 0.004478071350604296, + 0.009195003658533096, + 0.004784027114510536, + 0.018785757943987846, + 0.0106190275400877, + 0.010097228921949863, + 0.006790497340261936, + 0.014840289019048214, + 0.03502201661467552, + 0.009620755910873413, + 0.0116866584867239, + 0.037218041718006134, + 0.028150774538517, + 0.026990367099642754, + 0.022674469277262688, + 0.0350387804210186, + 0.028622623533010483, + 0.0407637394964695, + 0.028479333966970444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.003189953975379467 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255563884973526 + ], + "opposite_logprob_tokens": [ + -1.6005563735961914 + ], + "logprob_reward": -0.2255563884973526, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255563884973526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053700823336839676, + 0.04820407181978226, + 0.4306202530860901, + 0.024084489792585373, + 0.13968439400196075, + 0.024459240958094597, + 0.06584829837083817, + 0.039679013192653656, + 0.048222459852695465, + 0.125496968626976 + ], + "credit_map": [ + 0.007840795442461967, + 0.0030825594440102577, + 0.007049157749861479, + 0.0035274396650493145, + 0.003993359860032797, + 0.00618771743029356, + 0.006294384133070707, + 0.01572541333734989, + 0.005836864002048969, + 0.003060908056795597, + 0.00734324473887682, + 0.005798117257654667, + 0.006488830782473087, + 0.008796609938144684, + 0.004786328412592411, + 0.006093171890825033, + 0.1410793513059616, + 0.015422331169247627, + 0.01575872302055359, + 0.007975270040333271, + 0.02042224444448948, + 0.0315026119351387, + 0.026604587212204933, + 0.1718551367521286, + 0.0042664892971515656, + 0.0024813285563141108, + 0.002545767230913043, + 0.0018570361426100135, + 0.002473413711413741, + 0.0032641126308590174, + 0.0023236763663589954, + 0.004872667137533426, + 0.012647634372115135, + 0.01216441486030817, + 0.01927974261343479, + 0.03911999613046646, + 0.01387617364525795, + 0.007235195487737656, + 0.016949106007814407, + 0.01841212809085846, + 0.0032084924168884754, + 0.002851902274414897, + 0.003089230041950941, + 0.003883568337187171, + 0.0022574046161025763, + 0.003808454377576709, + 0.0030587324872612953, + 0.002301456406712532, + 0.008827424608170986, + 0.00896658655256033, + 0.008050776086747646, + 0.007072893902659416, + 0.00937274843454361, + 0.009615685790777206, + 0.007048605475574732, + 0.006893577519804239, + 0.005433017387986183, + 0.004612024873495102, + 0.008372724056243896, + 0.0046205902472138405, + 0.004509468097239733, + 0.0032133804634213448, + 0.005919191054999828, + 0.002998621203005314, + 0.007311163004487753, + 0.006271601654589176, + 0.005961562041193247, + 0.005388359539210796, + 0.007318383548408747, + 0.005152003839612007, + 0.005373466294258833, + 0.0054459222592413425, + 0.011131037026643753, + 0.020634138956665993, + 0.012681776657700539, + 0.013849830254912376, + 0.014219801872968674, + 0.016504278406500816, + 0.018039485439658165, + 0.018436623737215996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.2255563884973526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, artfully composed montage captures a teenage Pacific Islander girl twirling on a sunny beach, each moment a uniquely pictorial tableau. The sequence begins with a slow-motion medium shot, the camera gracefully arcing with her movements. She has long, wavy dark brown hair, a vibrant red hibiscus flower tucked behind her right ear. Her attire is a simple, lightweight sarong in a bright turquoise hue, its fabric billowing around her legs in the gentle sea breeze. As she twirls effortlessly on the damp, golden sand at the water's edge, the rapid cuts transition into a tight close-up of her face, a radiance of pure joy, her eyes crinkled in a genuine smile, her sun-kissed skin glowing warmly. The camera then tilts slightly downward as her bare feet kick up a fine mist of water and sand. The final shot in the series is a stunning wide shot, captured from a slightly lower angle, framing her mid-twirling, outstretched arms against the vast expanse of sparkling ocean and brilliant sunset with a breathtaking lens flare, bathing the entire scene in the warm, ethereal glow of the golden hour, evoking a powerful atmosphere of freedom, youth, and serene happiness.", + "seed": 19479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_11/index_7930_seed_19479_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": -0.01506555643863976, + "vlm_logprob": -0.01506555643863976, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.01506555643863976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015382226556539536 + ], + "opposite_logprob_tokens": [ + -4.765382289886475 + ], + "logprob_reward": -0.015382226556539536, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015382226556539536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12946636974811554, + 0.07205662876367569, + 0.05857834219932556, + 0.07693935930728912, + 0.10369307547807693, + 0.08198253810405731, + 0.09563380479812622, + 0.07380862534046173, + 0.061118777841329575, + 0.2467225193977356 + ], + "credit_map": [ + 0.017593706026673317, + 0.014758625999093056, + 0.023427464067935944, + 0.014321456663310528, + 0.012521468102931976, + 0.014869162812829018, + 0.011333438567817211, + 0.020641041919589043, + 0.00761808967217803, + 0.007502430584281683, + 0.009354793466627598, + 0.007744866423308849, + 0.007816844619810581, + 0.00974567886441946, + 0.012137524783611298, + 0.01013640034943819, + 0.0036253947764635086, + 0.005867558065801859, + 0.005773786921054125, + 0.004582079593092203, + 0.0087222158908844, + 0.012377170845866203, + 0.009079081006348133, + 0.008551054634153843, + 0.01028192974627018, + 0.01030502375215292, + 0.009273149073123932, + 0.006411269307136536, + 0.00875925924628973, + 0.010551784187555313, + 0.00861385278403759, + 0.012743092142045498, + 0.016468800604343414, + 0.009540154598653316, + 0.015824750065803528, + 0.008597325533628464, + 0.014558039605617523, + 0.0151956332847476, + 0.007550470065325499, + 0.015957895666360855, + 0.00960820447653532, + 0.00962152797728777, + 0.014427182264626026, + 0.006804489530622959, + 0.009644314646720886, + 0.012097224593162537, + 0.007316252216696739, + 0.012463347055017948, + 0.008533190935850143, + 0.006891417317092419, + 0.020905395969748497, + 0.006880518980324268, + 0.007419062778353691, + 0.015795959159731865, + 0.022188136354088783, + 0.0070201256312429905, + 0.007591185159981251, + 0.009444888681173325, + 0.007502611726522446, + 0.008581310510635376, + 0.010087945498526096, + 0.009885101579129696, + 0.009592870250344276, + 0.011122715659439564, + 0.013535168953239918, + 0.00687875971198082, + 0.00907512754201889, + 0.007480309344828129, + 0.0070048850029706955, + 0.006580424029380083, + 0.004837627988308668, + 0.00572647200897336, + 0.03478320688009262, + 0.0447722002863884, + 0.03617062419652939, + 0.05175614729523659, + 0.017944179475307465, + 0.013110712170600891, + 0.03382263705134392, + 0.014362807385623455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.015382226556539536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748858839273453 + ], + "opposite_logprob_tokens": [ + -5.006748676300049 + ], + "logprob_reward": -0.006748858839273453, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748858839273453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16485758125782013, + 0.07427451759576797, + 0.09363716840744019, + 0.07828444242477417, + 0.136423721909523, + 0.10876933485269547, + 0.15309107303619385, + 0.05033067986369133, + 0.04564415663480759, + 0.09468730539083481 + ], + "credit_map": [ + 0.026670297607779503, + 0.01461550872772932, + 0.035909343510866165, + 0.01707739569246769, + 0.01472562737762928, + 0.020283294841647148, + 0.012843886390328407, + 0.02273222804069519, + 0.0093240300193429, + 0.008356026373803616, + 0.008486366830766201, + 0.009049945510923862, + 0.008643053472042084, + 0.008758045732975006, + 0.010061472654342651, + 0.011595582589507103, + 0.007642423268407583, + 0.009823699481785297, + 0.012470574118196964, + 0.005783109925687313, + 0.012492459267377853, + 0.013465918600559235, + 0.02001366578042507, + 0.011945324949920177, + 0.009317804127931595, + 0.006966089829802513, + 0.006961741950362921, + 0.007482706569135189, + 0.010729936882853508, + 0.016896825283765793, + 0.009161864407360554, + 0.010767480358481407, + 0.027228517457842827, + 0.011412618681788445, + 0.023402230814099312, + 0.009152408689260483, + 0.01436679344624281, + 0.020960357040166855, + 0.010577651672065258, + 0.01932314969599247, + 0.012514716014266014, + 0.019015949219465256, + 0.014026946388185024, + 0.007467709947377443, + 0.01591789349913597, + 0.014593278989195824, + 0.009738446213304996, + 0.015494399704039097, + 0.014103710651397705, + 0.008593357168138027, + 0.031934335827827454, + 0.008395213633775711, + 0.010507027618587017, + 0.025162992998957634, + 0.044082578271627426, + 0.010311873629689217, + 0.006401302292943001, + 0.006072396412491798, + 0.005275401286780834, + 0.0049749077297747135, + 0.006539925932884216, + 0.007188646588474512, + 0.007672097999602556, + 0.0062060062773525715, + 0.007253664545714855, + 0.005722342059016228, + 0.006501726806163788, + 0.004856261424720287, + 0.005455272272229195, + 0.006203284487128258, + 0.003335776273161173, + 0.006315831560641527, + 0.009743285365402699, + 0.014320386573672295, + 0.014245716854929924, + 0.025412965565919876, + 0.005974949337542057, + 0.008241445757448673, + 0.00970164779573679, + 0.007046910468488932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.006748858839273453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009222545195370913 + ], + "opposite_logprob_tokens": [ + -7.000922203063965 + ], + "logprob_reward": -0.0009222545195370913, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009222545195370913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11016377061605453, + 0.12597422301769257, + 0.16378282010555267, + 0.10594118386507034, + 0.05768836662173271, + 0.047096312046051025, + 0.1012176126241684, + 0.07411336153745651, + 0.058303605765104294, + 0.15571871399879456 + ], + "credit_map": [ + 0.012454085983335972, + 0.009395265020430088, + 0.019741544499993324, + 0.014527239836752415, + 0.007848731242120266, + 0.012027692049741745, + 0.011817296035587788, + 0.022351915016770363, + 0.013502583838999271, + 0.013182880356907845, + 0.010986624285578728, + 0.011973242275416851, + 0.011386287398636341, + 0.017908630892634392, + 0.025047261267900467, + 0.021986709907650948, + 0.01715812459588051, + 0.013456611894071102, + 0.027146097272634506, + 0.010568564757704735, + 0.025524325668811798, + 0.028766218572854996, + 0.016095425933599472, + 0.02506745047867298, + 0.017056485638022423, + 0.008381720632314682, + 0.010315250605344772, + 0.007654751185327768, + 0.01244289055466652, + 0.01650877855718136, + 0.013628829270601273, + 0.019952470436692238, + 0.012560238130390644, + 0.003813759656623006, + 0.010960280895233154, + 0.0038942296523600817, + 0.005376520566642284, + 0.009668191894888878, + 0.004180493764579296, + 0.0072346520610153675, + 0.00586154218763113, + 0.005352001171559095, + 0.004763968288898468, + 0.0035122402478009462, + 0.0071025690995156765, + 0.007072078995406628, + 0.003424616064876318, + 0.010007299482822418, + 0.008299111388623714, + 0.005926546175032854, + 0.020971445366740227, + 0.005398075096309185, + 0.007375635672360659, + 0.017110200598835945, + 0.029249591752886772, + 0.0068870107643306255, + 0.008491369895637035, + 0.009032404981553555, + 0.00854933075606823, + 0.007917879149317741, + 0.010434729978442192, + 0.010999226942658424, + 0.00740246893838048, + 0.011285952292382717, + 0.006845903117209673, + 0.006158933043479919, + 0.008516526781022549, + 0.006504547316581011, + 0.006345510948449373, + 0.007122585084289312, + 0.008445490151643753, + 0.008364109322428703, + 0.01949918642640114, + 0.01993311382830143, + 0.021825049072504044, + 0.025269515812397003, + 0.01470289658755064, + 0.020601786673069, + 0.015798961743712425, + 0.018088217824697495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.0009222545195370913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418930571526289 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.01418930571526289, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418930571526289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10787799209356308, + 0.08623865246772766, + 0.08443698287010193, + 0.051123104989528656, + 0.05965890362858772, + 0.27806147933006287, + 0.03403782844543457, + 0.055220648646354675, + 0.08416719734668732, + 0.15917716920375824 + ], + "credit_map": [ + 0.0170432198792696, + 0.009859134443104267, + 0.017634397372603416, + 0.007880233228206635, + 0.009447907097637653, + 0.015704916790127754, + 0.010137098841369152, + 0.02017107419669628, + 0.013099715113639832, + 0.008631279692053795, + 0.01074161659926176, + 0.007574193645268679, + 0.008115575648844242, + 0.010880540125072002, + 0.01207058597356081, + 0.015125147998332977, + 0.0053968122228980064, + 0.009591305628418922, + 0.01080052275210619, + 0.006934007629752159, + 0.011186666786670685, + 0.015642978250980377, + 0.013709891587495804, + 0.011174800805747509, + 0.007510570343583822, + 0.0060715265572071075, + 0.005150671117007732, + 0.004118853248655796, + 0.0069565013982355595, + 0.009198090061545372, + 0.005001625511795282, + 0.007115271408110857, + 0.011096220463514328, + 0.004107754211872816, + 0.012096156366169453, + 0.0049797468818724155, + 0.004262997768819332, + 0.009674083441495895, + 0.006282724440097809, + 0.007159219589084387, + 0.03130367025732994, + 0.0443885400891304, + 0.029929179698228836, + 0.0204994585365057, + 0.04907780885696411, + 0.034370291978120804, + 0.029733022674918175, + 0.038759514689445496, + 0.0031249469611793756, + 0.0043620639480650425, + 0.0038313986733555794, + 0.0029509214218705893, + 0.004898793995380402, + 0.004700622987002134, + 0.00643354095518589, + 0.003735541831701994, + 0.006228184327483177, + 0.006940420251339674, + 0.007622790057212114, + 0.005256704520434141, + 0.007177318446338177, + 0.006731115281581879, + 0.00893619004637003, + 0.00632792292162776, + 0.006921837106347084, + 0.007004044950008392, + 0.012717492878437042, + 0.013257903978228569, + 0.00862585287541151, + 0.010115008801221848, + 0.013062713667750359, + 0.012462344020605087, + 0.020307330414652824, + 0.024461349472403526, + 0.019828226417303085, + 0.03809208795428276, + 0.012498883530497551, + 0.010047612711787224, + 0.02097334712743759, + 0.012968338094651699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.01418930571526289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808513656258583 + ], + "opposite_logprob_tokens": [ + -3.2880852222442627 + ], + "logprob_reward": -0.03808513656258583, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808513656258583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05018078163266182, + 0.07709912210702896, + 0.08892711997032166, + 0.06323728710412979, + 0.08397294580936432, + 0.042711853981018066, + 0.0729861631989479, + 0.05482751131057739, + 0.07093054056167603, + 0.39512667059898376 + ], + "credit_map": [ + 0.005906878039240837, + 0.005704971496015787, + 0.01124791894108057, + 0.004095135256648064, + 0.005447973031550646, + 0.005196923855692148, + 0.00497495848685503, + 0.007606018800288439, + 0.010300696827471256, + 0.006431717425584793, + 0.00813278742134571, + 0.007675162050873041, + 0.004431485664099455, + 0.011235344223678112, + 0.01099836453795433, + 0.017893552780151367, + 0.005766672547906637, + 0.006993403192609549, + 0.008211442269384861, + 0.005128988064825535, + 0.011866554617881775, + 0.01576833240687847, + 0.019567551091313362, + 0.015624182298779488, + 0.011681672185659409, + 0.006454950664192438, + 0.008801611140370369, + 0.0047232480719685555, + 0.006951594725251198, + 0.010243301279842854, + 0.006174881476908922, + 0.008206027559936047, + 0.017075130715966225, + 0.005141995381563902, + 0.014815381728112698, + 0.008338544517755508, + 0.011200537905097008, + 0.012125002220273018, + 0.005834429524838924, + 0.0094419214874506, + 0.005022021941840649, + 0.004497701767832041, + 0.006378093268722296, + 0.003697255626320839, + 0.00485884677618742, + 0.006055159028619528, + 0.005593509413301945, + 0.006609265226870775, + 0.007109195925295353, + 0.006296479608863592, + 0.011699315160512924, + 0.00435442803427577, + 0.005139591172337532, + 0.011701846495270729, + 0.02142953686416149, + 0.00525576900690794, + 0.006594761274755001, + 0.007767968345433474, + 0.007472602650523186, + 0.005191504489630461, + 0.007440917659550905, + 0.0085975993424654, + 0.004625860136002302, + 0.007136301137506962, + 0.013405416160821915, + 0.007218964863568544, + 0.0106089748442173, + 0.01026018988341093, + 0.005437417887151241, + 0.009298756718635559, + 0.005648331250995398, + 0.009052486158907413, + 0.052271150052547455, + 0.06662219017744064, + 0.055619627237319946, + 0.08299894630908966, + 0.031109625473618507, + 0.02560705505311489, + 0.05053667724132538, + 0.030361376702785492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.03808513656258583 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, artfully composed montage captures a teenage Pacific Islander girl twirling on a sunny beach, each moment a uniquely pictorial tableau. The sequence begins with a slow-motion medium shot, the camera gracefully arcing with her movements. She has long, wavy dark brown hair, a vibrant red hibiscus flower tucked behind her right ear. Her attire is a simple, lightweight sarong in a bright turquoise hue, its fabric billowing around her legs in the gentle sea breeze. As she twirls effortlessly on the damp, golden sand at the water's edge, the rapid cuts transition into a tight close-up of her face, a radiance of pure joy, her eyes crinkled in a genuine smile, her sun-kissed skin glowing warmly. The camera then tilts slightly downward as her bare feet kick up a fine mist of water and sand. The final shot in the series is a stunning wide shot, captured from a slightly lower angle, framing her mid-twirling, outstretched arms against the vast expanse of sparkling ocean and brilliant sunset with a breathtaking lens flare, bathing the entire scene in the warm, ethereal glow of the golden hour, evoking a powerful atmosphere of freedom, youth, and serene happiness.", + "seed": 19480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_11/index_7930_seed_19480_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": -0.04437596830539405, + "vlm_logprob": -0.04437596830539405, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.04437596830539405, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008266991935670376 + ], + "opposite_logprob_tokens": [ + -5.633266925811768 + ], + "logprob_reward": -0.008266991935670376, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008266991935670376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1899065524339676, + 0.1303303837776184, + 0.0567421056330204, + 0.026649175211787224, + 0.06731988489627838, + 0.14595963060855865, + 0.06758874654769897, + 0.10174299031496048, + 0.09353592246770859, + 0.12022460997104645 + ], + "credit_map": [ + 0.017206281423568726, + 0.026100218296051025, + 0.022811463102698326, + 0.05006055906414986, + 0.013308314606547356, + 0.02941410057246685, + 0.011219853535294533, + 0.019785765558481216, + 0.013555238954722881, + 0.008599869906902313, + 0.00787894893437624, + 0.030543023720383644, + 0.005494427867233753, + 0.012076347135007381, + 0.005281869322061539, + 0.04690065607428551, + 0.006455206777900457, + 0.006635678466409445, + 0.00773729057982564, + 0.0076665435917675495, + 0.0044801849871873856, + 0.007316598668694496, + 0.006991993170231581, + 0.009458606131374836, + 0.003756118705496192, + 0.0032392237335443497, + 0.00467144139111042, + 0.0041733114048838615, + 0.003610719460994005, + 0.0021271079313009977, + 0.0027692450676113367, + 0.0023020063526928425, + 0.006774914916604757, + 0.007052842527627945, + 0.012019870802760124, + 0.01456727646291256, + 0.008834104984998703, + 0.0041474588215351105, + 0.008895451202988625, + 0.005027958657592535, + 0.04042528197169304, + 0.013368681073188782, + 0.04550052061676979, + 0.016165591776371002, + 0.007163317874073982, + 0.006318259052932262, + 0.007420392706990242, + 0.009597585536539555, + 0.010485616512596607, + 0.004724833648651838, + 0.01847594976425171, + 0.009932669810950756, + 0.004419848322868347, + 0.008248353376984596, + 0.007582559250295162, + 0.0037189186550676823, + 0.014315938577055931, + 0.011732160113751888, + 0.0111829973757267, + 0.014605543576180935, + 0.01636023074388504, + 0.00854124128818512, + 0.011418771930038929, + 0.01358609925955534, + 0.007569109555333853, + 0.01596255786716938, + 0.011318176053464413, + 0.014848744496703148, + 0.007755774538964033, + 0.007591276429593563, + 0.009165233932435513, + 0.019325047731399536, + 0.0061479294672608376, + 0.017203211784362793, + 0.0070776864886283875, + 0.02157626673579216, + 0.006788602564483881, + 0.027056390419602394, + 0.010622195899486542, + 0.023752328008413315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.008266991935670376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265653133392334 + ], + "opposite_logprob_tokens": [ + -5.255265712738037 + ], + "logprob_reward": -0.005265653133392334, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265653133392334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15666529536247253, + 0.0953046903014183, + 0.03929295390844345, + 0.09385252743959427, + 0.07761551439762115, + 0.11646658927202225, + 0.06444304436445236, + 0.08835894614458084, + 0.10434974730014801, + 0.1636507362127304 + ], + "credit_map": [ + 0.01692577637732029, + 0.02131076343357563, + 0.01889975182712078, + 0.031443387269973755, + 0.014879179187119007, + 0.02196240983903408, + 0.01466023363173008, + 0.016583794727921486, + 0.008758625946938992, + 0.006232257001101971, + 0.009293967857956886, + 0.017173970118165016, + 0.006569609511643648, + 0.006815979722887278, + 0.014089704491198063, + 0.0263705812394619, + 0.0029081336688250303, + 0.0034302028361707926, + 0.004149740096181631, + 0.006020074710249901, + 0.005242591258138418, + 0.0039014234207570553, + 0.007619598414748907, + 0.0060211895033717155, + 0.012431482784450054, + 0.011512234807014465, + 0.018834508955478668, + 0.020756172016263008, + 0.008182031102478504, + 0.0073859659023582935, + 0.006905102636665106, + 0.00784502737224102, + 0.008424692787230015, + 0.010888480581343174, + 0.017222322523593903, + 0.013176478445529938, + 0.008473104797303677, + 0.006675749085843563, + 0.0076279654167592525, + 0.005126722157001495, + 0.029006781056523323, + 0.012289164587855339, + 0.03120996057987213, + 0.0176088809967041, + 0.005907908547669649, + 0.006075977347791195, + 0.007336522452533245, + 0.007031385321170092, + 0.0071454280987381935, + 0.00655157957226038, + 0.012749982997775078, + 0.013995747081935406, + 0.007444415241479874, + 0.003956580068916082, + 0.006384793668985367, + 0.006214516703039408, + 0.00959277804940939, + 0.011179526336491108, + 0.009447799995541573, + 0.012387903407216072, + 0.013685069978237152, + 0.006450133863836527, + 0.011053846217691898, + 0.014561890624463558, + 0.00927955936640501, + 0.01187286339700222, + 0.014716305769979954, + 0.010071085765957832, + 0.020876986905932426, + 0.015743518248200417, + 0.008918952196836472, + 0.012870481237769127, + 0.01009502075612545, + 0.029148660600185394, + 0.010755701921880245, + 0.029204189777374268, + 0.015082296915352345, + 0.02125426195561886, + 0.01592184416949749, + 0.032188769429922104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.005265653133392334 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193400101736188 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.003193400101736188, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193400101736188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.152167409658432, + 0.14696472883224487, + 0.10149600356817245, + 0.10349442809820175, + 0.13499298691749573, + 0.12327824532985687, + 0.05691160261631012, + 0.06804881989955902, + 0.039354342967271805, + 0.07329139858484268 + ], + "credit_map": [ + 0.01610223762691021, + 0.018577180802822113, + 0.020018775016069412, + 0.03751503303647041, + 0.011879895813763142, + 0.019289370626211166, + 0.01221610140055418, + 0.01656881347298622, + 0.012972746044397354, + 0.007826177403330803, + 0.008319337852299213, + 0.038041550666093826, + 0.003720139618963003, + 0.013310905545949936, + 0.006688710767775774, + 0.05608517304062843, + 0.014799109660089016, + 0.011962566524744034, + 0.01704518310725689, + 0.014567732810974121, + 0.00781883392482996, + 0.011622722260653973, + 0.009232727810740471, + 0.014447126537561417, + 0.011465614661574364, + 0.013720480725169182, + 0.016242532059550285, + 0.023078832775354385, + 0.010950573720037937, + 0.009457341395318508, + 0.007922658696770668, + 0.010656393133103848, + 0.012124322354793549, + 0.016549181193113327, + 0.02572440542280674, + 0.028171809390187263, + 0.018361354246735573, + 0.010128203779459, + 0.013715723529458046, + 0.01021798700094223, + 0.010365323163568974, + 0.008951883763074875, + 0.014097302220761776, + 0.016156794503331184, + 0.012133966200053692, + 0.018787020817399025, + 0.01547807827591896, + 0.02730787731707096, + 0.00806577317416668, + 0.0042541795410215855, + 0.013136805035173893, + 0.00776689825579524, + 0.006862209644168615, + 0.007918230257928371, + 0.005330564454197884, + 0.003576945746317506, + 0.010002676397562027, + 0.005512771662324667, + 0.017938682809472084, + 0.00746995909139514, + 0.006356037221848965, + 0.006860471796244383, + 0.00621653301641345, + 0.0076916893012821674, + 0.00317253265529871, + 0.006373830139636993, + 0.003615784924477339, + 0.005599516909569502, + 0.005624516401439905, + 0.0035328541416674852, + 0.004146616440266371, + 0.007288689259439707, + 0.004548899829387665, + 0.00960235670208931, + 0.0058700027875602245, + 0.010377729311585426, + 0.012300347909331322, + 0.011963934637606144, + 0.008570794947445393, + 0.010057336650788784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.003193400101736188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036139669828116894 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036139669828116894, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036139669828116894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1726834625005722, + 0.05422786995768547, + 0.040756482630968094, + 0.08237665146589279, + 0.08517637848854065, + 0.13803571462631226, + 0.10045413672924042, + 0.08849355578422546, + 0.08782560378313065, + 0.14997009932994843 + ], + "credit_map": [ + 0.023553285747766495, + 0.020639412105083466, + 0.026184825226664543, + 0.03258366882801056, + 0.014550898224115372, + 0.01900840364396572, + 0.013147301971912384, + 0.023015685379505157, + 0.005406116135418415, + 0.00623694434762001, + 0.009217469953000546, + 0.009844317100942135, + 0.006267333868891001, + 0.005266000982373953, + 0.00501513434574008, + 0.006974559277296066, + 0.007369136903434992, + 0.005205558612942696, + 0.007110989652574062, + 0.005230497568845749, + 0.003398963250219822, + 0.003498262260109186, + 0.0043264818377792835, + 0.0046165985986590385, + 0.01327487826347351, + 0.013915155082941055, + 0.011429408565163612, + 0.014226176775991917, + 0.007477254141122103, + 0.008354817517101765, + 0.006329851225018501, + 0.007369119208306074, + 0.00974117498844862, + 0.009231649339199066, + 0.017728179693222046, + 0.018546104431152344, + 0.008386634290218353, + 0.00794815644621849, + 0.008890445344150066, + 0.004704038146883249, + 0.03400269150733948, + 0.01398723665624857, + 0.03665676340460777, + 0.016879307106137276, + 0.0066501726396381855, + 0.012988205999135971, + 0.010319207794964314, + 0.0065521434880793095, + 0.014533228240907192, + 0.006932440213859081, + 0.02503317780792713, + 0.012438489124178886, + 0.009413053281605244, + 0.016150174662470818, + 0.010003015398979187, + 0.005950571969151497, + 0.013331674039363861, + 0.009015682153403759, + 0.0210263729095459, + 0.013152641244232655, + 0.005670215468853712, + 0.00810974556952715, + 0.007252284791320562, + 0.010934949852526188, + 0.008557080291211605, + 0.011903704144060612, + 0.010765853337943554, + 0.016423877328634262, + 0.009118671528995037, + 0.012246266938745975, + 0.008106432855129242, + 0.010703721083700657, + 0.011860073544085026, + 0.023556819185614586, + 0.016843751072883606, + 0.03174930810928345, + 0.015256915241479874, + 0.013044900260865688, + 0.022148849442601204, + 0.015509502030909061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0036139669828116894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a teenage Pacific Islander girl with long, wavy dark brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the girl wearing a bright turquoise sarong with a red hibiscus flower tucked behind her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the girl twirl on damp golden sand at the water's edge while barefoot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sequence include a slow-motion medium shot followed by a tight close-up of her smiling face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final shot show a wide view from a slightly lower angle with a lens flare against a sunset ocean?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153982937335968 + ], + "opposite_logprob_tokens": [ + -1.7015398740768433 + ], + "logprob_reward": -0.20153982937335968, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153982937335968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1037764921784401, + 0.13307711482048035, + 0.035701993852853775, + 0.08396020531654358, + 0.04728288948535919, + 0.11783801019191742, + 0.070367731153965, + 0.06033433973789215, + 0.060499101877212524, + 0.28716206550598145 + ], + "credit_map": [ + 0.01019713282585144, + 0.008579341694712639, + 0.01162259466946125, + 0.028316345065832138, + 0.009329420514404774, + 0.014457379467785358, + 0.006888755597174168, + 0.014385526068508625, + 0.01055599283427, + 0.008472524583339691, + 0.008697892539203167, + 0.034314267337322235, + 0.004300601780414581, + 0.011614174582064152, + 0.005564759485423565, + 0.049556903541088104, + 0.0038309397641569376, + 0.004597197286784649, + 0.005827585235238075, + 0.006224432494491339, + 0.0027244354132562876, + 0.0034390354994684458, + 0.003783364547416568, + 0.005275004077702761, + 0.01654093898832798, + 0.014097092673182487, + 0.011977139860391617, + 0.014362898655235767, + 0.006636528298258781, + 0.007543977815657854, + 0.004677617456763983, + 0.008124013431370258, + 0.004560628905892372, + 0.007513831369578838, + 0.008557356894016266, + 0.0068819476291537285, + 0.007266366388648748, + 0.0040846094489097595, + 0.00534584978595376, + 0.003072296967729926, + 0.03727375715970993, + 0.011255098506808281, + 0.039820171892642975, + 0.01101738028228283, + 0.004871868994086981, + 0.004117848817259073, + 0.0057164570316672325, + 0.0037654265761375427, + 0.009130235761404037, + 0.007009548135101795, + 0.014637872576713562, + 0.00788500253111124, + 0.009832873940467834, + 0.009064596146345139, + 0.007878976874053478, + 0.004928628448396921, + 0.009008221328258514, + 0.00503000570461154, + 0.01471000537276268, + 0.008047440089285374, + 0.004068521782755852, + 0.005682384129613638, + 0.005689890589565039, + 0.00809787679463625, + 0.004770201630890369, + 0.00963046308606863, + 0.006575404200702906, + 0.009457035921514034, + 0.006387295667082071, + 0.0072497115470469, + 0.010062972083687782, + 0.006366014014929533, + 0.025522788986563683, + 0.045054078102111816, + 0.029481632634997368, + 0.047536782920360565, + 0.01926453225314617, + 0.05339886620640755, + 0.023269163444638252, + 0.043634213507175446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.20153982937335968 + } + ], + "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_12/index_tensor([4427, 4427])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_12/index_tensor([4427, 4427])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ab239ed09750ddb106fd49d79ecc0a8293c156ce --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_12/index_tensor([4427, 4427])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This medium shot, captured from a slightly low angle with a handheld camera that subtly sways and undulates in sync with the water's movement, focuses on an elderly Asian man with thinning, silver-gray hair and a deeply wrinkled face. He sits in a small, weathered wooden boat, deliberately lowering his center of gravity and bending his knees to counteract a sudden numbness. His rough, gnarled hands, with prominent veins and calloused skin, instinctively grasp the gunwale of the dark, damp boat, his knuckles pale from the effort. The boat lurches on the unstable, dark waters, and the turbulent waves crash against the aged wood. The scene is illuminated by the dim, unfocused light of an overcast sky, creating a cool, desaturated palette of deep blues, grays, and dark browns that echo the damp boat. The man's expression is one of intense, quiet concentration, not fear, his gaze fixed on the turbulent waters ahead, creating a tense and somber atmosphere of profound and enduring resilience against the elements.", + "seed": 20778, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_12/index_4427_seed_20778_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": -0.03134238105267286, + "vlm_logprob": -0.03134238105267286, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03134238105267286, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035648226737976074 + ], + "opposite_logprob_tokens": [ + -3.4106483459472656 + ], + "logprob_reward": -0.035648226737976074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.035648226737976074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2671290934085846, + 0.06359758973121643, + 0.059632156044244766, + 0.07048090547323227, + 0.16281704604625702, + 0.21927683055400848, + 0.04060027003288269, + 0.03257007151842117, + 0.05467406287789345, + 0.029221994802355766 + ], + "credit_map": [ + 0.017671555280685425, + 0.033341184258461, + 0.02226901426911354, + 0.029642771929502487, + 0.04177563264966011, + 0.025042319670319557, + 0.05735855549573898, + 0.04002803936600685, + 0.0062732514925301075, + 0.008040833286941051, + 0.004875269252806902, + 0.00900465901941061, + 0.009963289834558964, + 0.006860196590423584, + 0.010323947295546532, + 0.008256140165030956, + 0.0061769792810082436, + 0.009036414325237274, + 0.004316987935453653, + 0.005073766689747572, + 0.014114949852228165, + 0.008940326981246471, + 0.0048467000015079975, + 0.0071260216645896435, + 0.005218758713454008, + 0.003984859213232994, + 0.0088118901476264, + 0.010058242827653885, + 0.007872241549193859, + 0.003968807402998209, + 0.017300935462117195, + 0.013265167362987995, + 0.015324401669204235, + 0.015041944570839405, + 0.01612873561680317, + 0.018712736666202545, + 0.03318279981613159, + 0.03151128813624382, + 0.014546102844178677, + 0.018369035795331, + 0.02021965943276882, + 0.022959213703870773, + 0.017889920622110367, + 0.027522293850779533, + 0.04177238419651985, + 0.041229646652936935, + 0.025130506604909897, + 0.02255319058895111, + 0.004182416945695877, + 0.0032965238206088543, + 0.004832680802792311, + 0.003072880208492279, + 0.005620765965431929, + 0.008013658225536346, + 0.004045965149998665, + 0.007535374257713556, + 0.0032248608767986298, + 0.003147309413179755, + 0.0025980444625020027, + 0.0033632926642894745, + 0.0035921407397836447, + 0.00642079534009099, + 0.00565042719244957, + 0.004573197104036808, + 0.00670073414221406, + 0.007384219206869602, + 0.008300299756228924, + 0.006069740280508995, + 0.005984599236398935, + 0.004939773119986057, + 0.007982143200933933, + 0.007312547415494919, + 0.0028429690282791853, + 0.0024408039171248674, + 0.004768128041177988, + 0.0026478443760424852, + 0.004237900488078594, + 0.0033256004098802805, + 0.003958205692470074, + 0.005000542849302292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.035648226737976074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649660550057888 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649660550057888, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649660550057888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29243040084838867, + 0.05679084360599518, + 0.04166475310921669, + 0.12352185696363449, + 0.0694659873843193, + 0.23894478380680084, + 0.033609822392463684, + 0.052599236369132996, + 0.06893126666545868, + 0.02204108238220215 + ], + "credit_map": [ + 0.03164668381214142, + 0.040401097387075424, + 0.037499330937862396, + 0.03194594383239746, + 0.02512887679040432, + 0.025141704827547073, + 0.05900596082210541, + 0.04166080430150032, + 0.006082087755203247, + 0.00645421864464879, + 0.005564202554523945, + 0.00815731380134821, + 0.007769434712827206, + 0.005525865592062473, + 0.009214414283633232, + 0.008023305796086788, + 0.00511253671720624, + 0.0033482445869594812, + 0.00299066212028265, + 0.004124908242374659, + 0.0071705421432852745, + 0.0036612479016184807, + 0.005725311115384102, + 0.009531303308904171, + 0.017101407051086426, + 0.007340375799685717, + 0.011995941400527954, + 0.010625793598592281, + 0.014001703821122646, + 0.007698013447225094, + 0.016615672037005424, + 0.03814295306801796, + 0.006578968837857246, + 0.007390118204057217, + 0.006350026465952396, + 0.008171577006578445, + 0.014229526743292809, + 0.012897323817014694, + 0.006078027654439211, + 0.007770413998514414, + 0.022852659225463867, + 0.02858772687613964, + 0.021023347973823547, + 0.03064809739589691, + 0.04366561770439148, + 0.04222782328724861, + 0.025409836322069168, + 0.02452966570854187, + 0.00508326618000865, + 0.0028781022410839796, + 0.0036767658311873674, + 0.003640411188825965, + 0.003613861510530114, + 0.004616094287484884, + 0.0033293121960014105, + 0.006772007327526808, + 0.005372567102313042, + 0.004369210451841354, + 0.007958903908729553, + 0.00899645034223795, + 0.004525784403085709, + 0.006274604704231024, + 0.007449860218912363, + 0.0076518552377820015, + 0.006171520799398422, + 0.0073423925787210464, + 0.007411769591271877, + 0.006737933494150639, + 0.010921199806034565, + 0.004114959854632616, + 0.007786534260958433, + 0.018444959074258804, + 0.0020817690528929234, + 0.0028504952788352966, + 0.004496510606259108, + 0.0018878265982493758, + 0.003406772157177329, + 0.001485878019593656, + 0.00294515467248857, + 0.002886676462367177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.004649660550057888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897964119911194 + ], + "opposite_logprob_tokens": [ + -2.578979730606079 + ], + "logprob_reward": -0.07897964119911194, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897964119911194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30455419421195984, + 0.051251817494630814, + 0.0636814683675766, + 0.05219126120209694, + 0.07700914889574051, + 0.07323457300662994, + 0.035873983055353165, + 0.2072058916091919, + 0.08127575367689133, + 0.05372191220521927 + ], + "credit_map": [ + 0.030103333294391632, + 0.03769618272781372, + 0.041466034948825836, + 0.033045146614313126, + 0.02523760311305523, + 0.0238818246871233, + 0.06651686131954193, + 0.046607211232185364, + 0.005209752824157476, + 0.005345471203327179, + 0.0047590285539627075, + 0.00792065542191267, + 0.006230351980775595, + 0.005626225378364325, + 0.009472967125475407, + 0.0066873603500425816, + 0.0066415960900485516, + 0.006535694468766451, + 0.006477897521108389, + 0.0070549906231462955, + 0.009927568025887012, + 0.005998851265758276, + 0.011720344424247742, + 0.009324528276920319, + 0.00749923987314105, + 0.005346369929611683, + 0.0038566458970308304, + 0.003234144067391753, + 0.007509761489927769, + 0.004497025161981583, + 0.00648408243432641, + 0.013763993978500366, + 0.008226477541029453, + 0.006417013704776764, + 0.007681759539991617, + 0.007640708703547716, + 0.01328301802277565, + 0.01718616485595703, + 0.006521112751215696, + 0.010052889585494995, + 0.006898554041981697, + 0.011828050948679447, + 0.012606652453541756, + 0.009007754735648632, + 0.007730548270046711, + 0.006361690349876881, + 0.009862685576081276, + 0.008938632905483246, + 0.0035171399358659983, + 0.003015296533703804, + 0.0043020714074373245, + 0.004659208934754133, + 0.0041787466034293175, + 0.00519143370911479, + 0.004354802425950766, + 0.006655283737927675, + 0.017457880079746246, + 0.018381398171186447, + 0.01534679438918829, + 0.034467730671167374, + 0.022255178540945053, + 0.030531220138072968, + 0.039206985384225845, + 0.029558716341853142, + 0.00750098517164588, + 0.008348199538886547, + 0.01086802314966917, + 0.0069086854346096516, + 0.012742347083985806, + 0.006437218282371759, + 0.010819477029144764, + 0.017650814726948738, + 0.006873033940792084, + 0.010745364241302013, + 0.00830815639346838, + 0.0034884950146079063, + 0.007214668206870556, + 0.008442101068794727, + 0.004119101446121931, + 0.0045309909619390965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.07897964119911194 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636152133345604 + ], + "opposite_logprob_tokens": [ + -3.6513614654541016 + ], + "logprob_reward": -0.02636152133345604, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636152133345604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.357862263917923, + 0.06358563154935837, + 0.019521348178386688, + 0.021675750613212585, + 0.05637389048933983, + 0.372389554977417, + 0.011184374801814556, + 0.02032904140651226, + 0.05305914953351021, + 0.024018943309783936 + ], + "credit_map": [ + 0.04129188135266304, + 0.05171043053269386, + 0.047045618295669556, + 0.03653213381767273, + 0.0327138788998127, + 0.03423719480633736, + 0.06527572125196457, + 0.04905546456575394, + 0.006390243303030729, + 0.008146260865032673, + 0.005844212137162685, + 0.0076385727152228355, + 0.008492874912917614, + 0.007432769052684307, + 0.010439741425216198, + 0.00920096319168806, + 0.0018720935331657529, + 0.0029751264955848455, + 0.001832216395996511, + 0.002048643771559, + 0.003015791764482856, + 0.0016911836573854089, + 0.0029872392769902945, + 0.003099056426435709, + 0.0018626475939527154, + 0.003509214846417308, + 0.001573895220644772, + 0.0027017060201615095, + 0.0031519162002950907, + 0.0016736534889787436, + 0.002908462891355157, + 0.004294256214052439, + 0.005632488057017326, + 0.00697874603793025, + 0.005234966985881329, + 0.00634271465241909, + 0.011749240569770336, + 0.009533087722957134, + 0.0048266989178955555, + 0.006075948942452669, + 0.03537136688828468, + 0.04707702249288559, + 0.035898707807064056, + 0.04889528825879097, + 0.06567195057868958, + 0.062147337943315506, + 0.038153938949108124, + 0.039173975586891174, + 0.0013078959891572595, + 0.001124994596466422, + 0.001351526123471558, + 0.0015528915682807565, + 0.001340103684924543, + 0.0017162234289571643, + 0.0011725438525900245, + 0.0016181976534426212, + 0.0018181460909545422, + 0.0019479964394122362, + 0.002480648225173354, + 0.002161642536520958, + 0.0020994110964238644, + 0.0027869190089404583, + 0.004137134179472923, + 0.002897145226597786, + 0.004785131197422743, + 0.004558860789984465, + 0.006577680818736553, + 0.0043365228921175, + 0.008722096681594849, + 0.004626804497092962, + 0.005276041571050882, + 0.014176013879477978, + 0.002447140868753195, + 0.0032813868019729853, + 0.0032769974786788225, + 0.0020038781221956015, + 0.0035185106098651886, + 0.003715396858751774, + 0.0030570696108043194, + 0.002718564122915268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02636152133345604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072855442762375 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072855442762375, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072855442762375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19039179384708405, + 0.09330949932336807, + 0.07205626368522644, + 0.06548323482275009, + 0.12430895864963531, + 0.2763703763484955, + 0.03782893344759941, + 0.04330504313111305, + 0.05561962351202965, + 0.04132635518908501 + ], + "credit_map": [ + 0.02262498252093792, + 0.03268169239163399, + 0.013824204914271832, + 0.016367554664611816, + 0.03109123930335045, + 0.01740359701216221, + 0.03107895329594612, + 0.025319576263427734, + 0.010610698722302914, + 0.013126485981047153, + 0.007168663665652275, + 0.01242570299655199, + 0.012914905324578285, + 0.00995364598929882, + 0.013811314478516579, + 0.013298086822032928, + 0.008406060747802258, + 0.010052858851850033, + 0.0064805978909134865, + 0.007491227239370346, + 0.013612532056868076, + 0.006941421423107386, + 0.008953562006354332, + 0.010118002071976662, + 0.004473687615245581, + 0.005142400972545147, + 0.00610879622399807, + 0.009123536758124828, + 0.009388871490955353, + 0.005751174408942461, + 0.01619378663599491, + 0.009300983510911465, + 0.014225364662706852, + 0.01993350125849247, + 0.011433652602136135, + 0.012621539644896984, + 0.024849649518728256, + 0.018245195969939232, + 0.008330065757036209, + 0.01466997992247343, + 0.026354385539889336, + 0.025927739217877388, + 0.01925288699567318, + 0.031781770288944244, + 0.05738487094640732, + 0.054165955632925034, + 0.03398803621530533, + 0.027514733374118805, + 0.0055862730368971825, + 0.003010509302839637, + 0.004847222473472357, + 0.004079952836036682, + 0.004290178418159485, + 0.005509485956281424, + 0.003420541062951088, + 0.0070847705937922, + 0.00356394168920815, + 0.003609331790357828, + 0.006818052846938372, + 0.0065470049157738686, + 0.0034025898203253746, + 0.005205436144024134, + 0.00696286465972662, + 0.007195817772299051, + 0.005467822775244713, + 0.004155727103352547, + 0.005967775825411081, + 0.005057758651673794, + 0.00857275165617466, + 0.00681238854303956, + 0.005725303199142218, + 0.013860099017620087, + 0.0033028002362698317, + 0.004286875948309898, + 0.005694772582501173, + 0.005215606652200222, + 0.006202560383826494, + 0.004101809114217758, + 0.008533583022654057, + 0.003988342825323343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.011072855442762375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium shot, captured from a slightly low angle with a handheld camera that subtly sways and undulates in sync with the water's movement, focuses on an elderly Asian man with thinning, silver-gray hair and a deeply wrinkled face. He sits in a small, weathered wooden boat, deliberately lowering his center of gravity and bending his knees to counteract a sudden numbness. His rough, gnarled hands, with prominent veins and calloused skin, instinctively grasp the gunwale of the dark, damp boat, his knuckles pale from the effort. The boat lurches on the unstable, dark waters, and the turbulent waves crash against the aged wood. The scene is illuminated by the dim, unfocused light of an overcast sky, creating a cool, desaturated palette of deep blues, grays, and dark browns that echo the damp boat. The man's expression is one of intense, quiet concentration, not fear, his gaze fixed on the turbulent waters ahead, creating a tense and somber atmosphere of profound and enduring resilience against the elements.", + "seed": 20779, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_12/index_4427_seed_20779_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": -0.017740753944963217, + "vlm_logprob": -0.017740753944963217, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.017740753944963217, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010135963559150696 + ], + "opposite_logprob_tokens": [ + -4.885136127471924 + ], + "logprob_reward": -0.010135963559150696, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010135963559150696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13182300329208374, + 0.0960177406668663, + 0.0905698612332344, + 0.02724103443324566, + 0.04118958115577698, + 0.06136854365468025, + 0.06798641383647919, + 0.14447830617427826, + 0.19925081729888916, + 0.14007475972175598 + ], + "credit_map": [ + 0.014177389442920685, + 0.02847941964864731, + 0.0183738861232996, + 0.018881238996982574, + 0.01533144898712635, + 0.008645785972476006, + 0.012559034861624241, + 0.015374805778265, + 0.0074784643948078156, + 0.01734745129942894, + 0.007695219479501247, + 0.012988249771296978, + 0.01360575295984745, + 0.010972143150866032, + 0.018264226615428925, + 0.007666231598705053, + 0.012510309927165508, + 0.008670290932059288, + 0.009205606766045094, + 0.01095744501799345, + 0.016830939799547195, + 0.012129678390920162, + 0.012042952701449394, + 0.00822264701128006, + 0.0029890495352447033, + 0.00354744796641171, + 0.004471666645258665, + 0.003505901899188757, + 0.00418937299400568, + 0.0032180841080844402, + 0.0032902457751333714, + 0.0020292645785957575, + 0.004074258729815483, + 0.005923275835812092, + 0.004172957967966795, + 0.008841682225465775, + 0.005402786657214165, + 0.003961509559303522, + 0.00519738532602787, + 0.003615726949647069, + 0.006677774712443352, + 0.010776689276099205, + 0.00752382492646575, + 0.01040684711188078, + 0.005984525661915541, + 0.005356751848012209, + 0.009072010405361652, + 0.005570117849856615, + 0.01210491918027401, + 0.011160911060869694, + 0.0059698219411075115, + 0.008281385526061058, + 0.004510119091719389, + 0.003858191194012761, + 0.014346130192279816, + 0.00775492750108242, + 0.018371807411313057, + 0.024927355349063873, + 0.019564958289265633, + 0.015106989070773125, + 0.014803511090576649, + 0.013875934295356274, + 0.02556530199944973, + 0.012262442149221897, + 0.01550426334142685, + 0.03479849547147751, + 0.024515807628631592, + 0.024195801466703415, + 0.04103543609380722, + 0.015715794637799263, + 0.02780870720744133, + 0.01567651890218258, + 0.006285378709435463, + 0.017236342653632164, + 0.013201533816754818, + 0.01817413792014122, + 0.041016172617673874, + 0.01542361918836832, + 0.017637021839618683, + 0.01110054925084114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.010135963559150696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041035739704966545 + ], + "opposite_logprob_tokens": [ + -5.504103660583496 + ], + "logprob_reward": -0.0041035739704966545, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041035739704966545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07048120349645615, + 0.06318322569131851, + 0.21236909925937653, + 0.034810371696949005, + 0.03774179145693779, + 0.06365437805652618, + 0.2429407387971878, + 0.1627122163772583, + 0.07326557487249374, + 0.03884143754839897 + ], + "credit_map": [ + 0.006623607594519854, + 0.013101018965244293, + 0.007246152963489294, + 0.01116377767175436, + 0.00785823818296194, + 0.007054143585264683, + 0.008219631388783455, + 0.009214631281793118, + 0.016418399289250374, + 0.006876885425299406, + 0.0027734546456485987, + 0.005996025167405605, + 0.004491103813052177, + 0.015309025533497334, + 0.007511071395128965, + 0.0038072592578828335, + 0.043093886226415634, + 0.03799446299672127, + 0.02901095151901245, + 0.01643531396985054, + 0.02843288518488407, + 0.02818075567483902, + 0.01741485856473446, + 0.011805977672338486, + 0.003993576858192682, + 0.0027171492110937834, + 0.008312657475471497, + 0.003413013182580471, + 0.0059732478111982346, + 0.004290154203772545, + 0.0034633360337466, + 0.0026472387835383415, + 0.004044630564749241, + 0.004690307192504406, + 0.004774533212184906, + 0.007742949295789003, + 0.005607303697615862, + 0.002447191160172224, + 0.0050108772702515125, + 0.0034239981323480606, + 0.006566742900758982, + 0.010243939235806465, + 0.009466193616390228, + 0.011103659868240356, + 0.0064587970264256, + 0.005929657258093357, + 0.009701781906187534, + 0.004183611366897821, + 0.023453250527381897, + 0.038855716586112976, + 0.049788426607847214, + 0.02531559392809868, + 0.014885397627949715, + 0.015763333067297935, + 0.027655594050884247, + 0.04722343385219574, + 0.01254063006490469, + 0.023242997005581856, + 0.01950264349579811, + 0.027109062299132347, + 0.014299777336418629, + 0.01152687706053257, + 0.035220514982938766, + 0.019269701093435287, + 0.006131547968834639, + 0.012346378527581692, + 0.00886852853000164, + 0.008521636947989464, + 0.016245439648628235, + 0.005711992736905813, + 0.009316693060100079, + 0.006123358383774757, + 0.00275948503986001, + 0.006152905523777008, + 0.005719021428376436, + 0.005688128527253866, + 0.0065404195338487625, + 0.0030454455409199, + 0.005078752059489489, + 0.0038572798948735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.0041035739704966545 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048652052879333496 + ], + "opposite_logprob_tokens": [ + -3.048652172088623 + ], + "logprob_reward": -0.048652052879333496, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048652052879333496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0778343454003334, + 0.12036802619695663, + 0.16296526789665222, + 0.12083180248737335, + 0.06003667786717415, + 0.025371689349412918, + 0.09034600108861923, + 0.07257559895515442, + 0.22305138409137726, + 0.04661915451288223 + ], + "credit_map": [ + 0.007242335006594658, + 0.014641094952821732, + 0.00875162798911333, + 0.01046174205839634, + 0.007705862168222666, + 0.006371198687702417, + 0.009299908764660358, + 0.013360581360757351, + 0.0352490209043026, + 0.011473209597170353, + 0.004619037266820669, + 0.008586499840021133, + 0.008261657319962978, + 0.032108381390571594, + 0.010410485789179802, + 0.009659749455749989, + 0.04286418482661247, + 0.03165583685040474, + 0.021875349804759026, + 0.012565226294100285, + 0.016006212681531906, + 0.01899460330605507, + 0.010895675979554653, + 0.008108198642730713, + 0.01426115445792675, + 0.013263121247291565, + 0.012568309903144836, + 0.013878827914595604, + 0.02438296563923359, + 0.017120275646448135, + 0.0147639699280262, + 0.010593190789222717, + 0.004850238095968962, + 0.008018250577151775, + 0.0043163001537323, + 0.011045929975807667, + 0.010565420612692833, + 0.0053749000653624535, + 0.007980898022651672, + 0.007884746417403221, + 0.0020435412880033255, + 0.004249224439263344, + 0.002861168934032321, + 0.005086236167699099, + 0.0031210132874548435, + 0.002178701339289546, + 0.002615793840959668, + 0.0032160095870494843, + 0.011897332966327667, + 0.014852135442197323, + 0.01259569451212883, + 0.011708628386259079, + 0.006940584629774094, + 0.005479205399751663, + 0.016896704211831093, + 0.009975717402994633, + 0.012386049143970013, + 0.010950223542749882, + 0.013139679096639156, + 0.008231553249061108, + 0.006740638520568609, + 0.00486408919095993, + 0.01104513369500637, + 0.005218238569796085, + 0.01766596920788288, + 0.03721218183636665, + 0.025932172313332558, + 0.02457978017628193, + 0.05270661786198616, + 0.02069753408432007, + 0.029136067256331444, + 0.015121068805456161, + 0.0023909718729555607, + 0.006223710719496012, + 0.003128092736005783, + 0.006472434848546982, + 0.014391676522791386, + 0.004381334409117699, + 0.005806935951113701, + 0.0038240018766373396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.048652052879333496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11815904825925827, + 0.09186378121376038, + 0.19649992883205414, + 0.035766929388046265, + 0.0760749876499176, + 0.12325719743967056, + 0.06190894916653633, + 0.20242106914520264, + 0.03484008088707924, + 0.05920812487602234 + ], + "credit_map": [ + 0.010054579935967922, + 0.02660396136343479, + 0.013449481688439846, + 0.01634492352604866, + 0.013670478016138077, + 0.009321524761617184, + 0.012090741656720638, + 0.016623342409729958, + 0.011776181869208813, + 0.017309952527284622, + 0.006993326358497143, + 0.010701831430196762, + 0.008866995573043823, + 0.013564329594373703, + 0.01311060693114996, + 0.009540541097521782, + 0.024094469845294952, + 0.02611960843205452, + 0.029911069199442863, + 0.012847996316850185, + 0.03247551620006561, + 0.034904755651950836, + 0.01806173287332058, + 0.01808476448059082, + 0.003993378020823002, + 0.0043637449853122234, + 0.005122337955981493, + 0.0037815142422914505, + 0.005556938704103231, + 0.005502635147422552, + 0.004437355324625969, + 0.0030090215150266886, + 0.00963667407631874, + 0.010184579528868198, + 0.007720292080193758, + 0.014871959574520588, + 0.010715201497077942, + 0.004982426762580872, + 0.00981283187866211, + 0.008151012472808361, + 0.0169809740036726, + 0.021787486970424652, + 0.018812555819749832, + 0.020225077867507935, + 0.009504975751042366, + 0.009693781845271587, + 0.01598379760980606, + 0.010268543846905231, + 0.011265137232840061, + 0.010178360156714916, + 0.011135057546198368, + 0.007194872014224529, + 0.0039613088592886925, + 0.005899201612919569, + 0.006746275350451469, + 0.005528723355382681, + 0.032493989914655685, + 0.027908755466341972, + 0.03392426669597626, + 0.032544106245040894, + 0.016552118584513664, + 0.01203357893973589, + 0.02923455461859703, + 0.017729684710502625, + 0.00329730031080544, + 0.005536565091460943, + 0.0034924752544611692, + 0.004353020805865526, + 0.006042862311005592, + 0.003111777361482382, + 0.0063408962450921535, + 0.002665181178599596, + 0.003898083232343197, + 0.008531483821570873, + 0.007822845131158829, + 0.00849666353315115, + 0.013785271905362606, + 0.0038886340335011482, + 0.008265323005616665, + 0.004519812762737274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.007631078828126192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181100487709045 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.018181100487709045, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181100487709045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19957976043224335, + 0.07229994237422943, + 0.2083086520433426, + 0.06659714877605438, + 0.0632820650935173, + 0.04311157763004303, + 0.10972508788108826, + 0.06468819826841354, + 0.04980562999844551, + 0.1226019635796547 + ], + "credit_map": [ + 0.015059654600918293, + 0.046591565012931824, + 0.016359882429242134, + 0.03007691167294979, + 0.02231251448392868, + 0.024309393018484116, + 0.018386010080575943, + 0.026483828201889992, + 0.008318164385855198, + 0.014887970872223377, + 0.0040152654983103275, + 0.008923931047320366, + 0.01127605140209198, + 0.010104709304869175, + 0.008678771555423737, + 0.006095079239457846, + 0.050657425075769424, + 0.02492513693869114, + 0.03681468963623047, + 0.016050344333052635, + 0.02441655844449997, + 0.022441748529672623, + 0.016602160409092903, + 0.01640057936310768, + 0.006856151390820742, + 0.00951840914785862, + 0.008550918661057949, + 0.005191924050450325, + 0.011319280602037907, + 0.012232634238898754, + 0.0065526035614311695, + 0.006375226657837629, + 0.006710460875183344, + 0.007734380196779966, + 0.004859001375734806, + 0.010822596959769726, + 0.006976246368139982, + 0.005709243472665548, + 0.006984518375247717, + 0.013485622592270374, + 0.004477516748011112, + 0.00652918079867959, + 0.004091409035027027, + 0.00775564881041646, + 0.004226934164762497, + 0.00528945866972208, + 0.005950594320893288, + 0.004790835082530975, + 0.023826584219932556, + 0.01882115565240383, + 0.018100036308169365, + 0.010270744562149048, + 0.007580186706036329, + 0.009167324751615524, + 0.012440294958651066, + 0.009518753737211227, + 0.007134830113500357, + 0.01137937605381012, + 0.00994435977190733, + 0.007829925045371056, + 0.006632796488702297, + 0.004022765439003706, + 0.011996607296168804, + 0.005747534334659576, + 0.008454037830233574, + 0.00796426273882389, + 0.0066437507048249245, + 0.0042846594005823135, + 0.007721956353634596, + 0.004462384153157473, + 0.005253509618341923, + 0.005021072458475828, + 0.006129818968474865, + 0.015589521266520023, + 0.009450830519199371, + 0.015621974132955074, + 0.034007780253887177, + 0.013376678340137005, + 0.01739482954144478, + 0.011030521243810654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.018181100487709045 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium shot, captured from a slightly low angle with a handheld camera that subtly sways and undulates in sync with the water's movement, focuses on an elderly Asian man with thinning, silver-gray hair and a deeply wrinkled face. He sits in a small, weathered wooden boat, deliberately lowering his center of gravity and bending his knees to counteract a sudden numbness. His rough, gnarled hands, with prominent veins and calloused skin, instinctively grasp the gunwale of the dark, damp boat, his knuckles pale from the effort. The boat lurches on the unstable, dark waters, and the turbulent waves crash against the aged wood. The scene is illuminated by the dim, unfocused light of an overcast sky, creating a cool, desaturated palette of deep blues, grays, and dark browns that echo the damp boat. The man's expression is one of intense, quiet concentration, not fear, his gaze fixed on the turbulent waters ahead, creating a tense and somber atmosphere of profound and enduring resilience against the elements.", + "seed": 20780, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_12/index_4427_seed_20780_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": -0.08088875627145171, + "vlm_logprob": -0.08088875627145171, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.08088875627145171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0261221956461668 + ], + "opposite_logprob_tokens": [ + -3.7761220932006836 + ], + "logprob_reward": -0.0261221956461668, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0261221956461668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13386401534080505, + 0.05878835916519165, + 0.05459222197532654, + 0.09107274562120438, + 0.037160880863666534, + 0.016582384705543518, + 0.38618385791778564, + 0.08399252593517303, + 0.06937268376350403, + 0.06839033216238022 + ], + "credit_map": [ + 0.051369696855545044, + 0.05398954823613167, + 0.0031050778925418854, + 0.0031575236935168505, + 0.007011157926172018, + 0.008322632871568203, + 0.0032199060078710318, + 0.00368846976198256, + 0.007409829180687666, + 0.00946841947734356, + 0.005547249224036932, + 0.0060768635012209415, + 0.011492442339658737, + 0.005947411525994539, + 0.005520657170563936, + 0.0073254830203950405, + 0.006486811209470034, + 0.008160640485584736, + 0.004632037598639727, + 0.0050192526541650295, + 0.011281277984380722, + 0.005719724576920271, + 0.005007871892303228, + 0.008284601382911205, + 0.006925617344677448, + 0.0076577989384531975, + 0.0067207831889390945, + 0.008671029470860958, + 0.01541926059871912, + 0.012518721632659435, + 0.01970037817955017, + 0.01345914788544178, + 0.003145435359328985, + 0.0031449845992028713, + 0.002520873211324215, + 0.006994624622166157, + 0.0033099791035056114, + 0.003133926074951887, + 0.010494918562471867, + 0.004416139330714941, + 0.0019407885847613215, + 0.00271755107678473, + 0.0018544585909694433, + 0.002560982946306467, + 0.0017645281041041017, + 0.0024417999666184187, + 0.0016020378097891808, + 0.0017002378590404987, + 0.020643116906285286, + 0.023952849209308624, + 0.03373027592897415, + 0.021467536687850952, + 0.025216957554221153, + 0.05760464072227478, + 0.07626321911811829, + 0.12730525434017181, + 0.006166430190205574, + 0.013082117773592472, + 0.01137013453990221, + 0.012035873718559742, + 0.01066980604082346, + 0.014382917433977127, + 0.008928511291742325, + 0.007356724236160517, + 0.004613560624420643, + 0.008250297047197819, + 0.009040883742272854, + 0.009940066374838352, + 0.008982262574136257, + 0.007076420355588198, + 0.013128970749676228, + 0.00834022555500269, + 0.006045532878488302, + 0.009943533688783646, + 0.005619596689939499, + 0.012503055855631828, + 0.008972340263426304, + 0.006795524153858423, + 0.00698089599609375, + 0.011529841460287571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0261221956461668 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764006981626153 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.00764006981626153, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764006981626153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059346918016672134, + 0.09763754159212112, + 0.07202105969190598, + 0.13148829340934753, + 0.06071272864937782, + 0.042067401111125946, + 0.34431278705596924, + 0.045842595398426056, + 0.053440701216459274, + 0.0931299701333046 + ], + "credit_map": [ + 0.0154058076441288, + 0.01763315685093403, + 0.0039010243490338326, + 0.003030354157090187, + 0.005875412840396166, + 0.007323016878217459, + 0.0028769862838089466, + 0.0033011536579579115, + 0.012975463643670082, + 0.013737188652157784, + 0.009399473667144775, + 0.011542941443622112, + 0.01680278405547142, + 0.011026989668607712, + 0.009112287312746048, + 0.013040419667959213, + 0.007286887615919113, + 0.010012178681790829, + 0.005124215502291918, + 0.008384809829294682, + 0.013581548817455769, + 0.008167642168700695, + 0.008314820006489754, + 0.011148955672979355, + 0.015087824314832687, + 0.013972158543765545, + 0.012420335784554482, + 0.018235402181744576, + 0.01816382445394993, + 0.017159778624773026, + 0.015505696646869183, + 0.020943285897374153, + 0.004908097442239523, + 0.00694365706294775, + 0.0034790418576449156, + 0.011468937620520592, + 0.005377958994358778, + 0.004224639385938644, + 0.018012920394539833, + 0.006297477055341005, + 0.003991540055721998, + 0.0068573253229260445, + 0.0035867278929799795, + 0.003916229121387005, + 0.004944020416587591, + 0.007358290255069733, + 0.004391090478748083, + 0.007022174075245857, + 0.018693730235099792, + 0.021853288635611534, + 0.02700149081647396, + 0.019026894122362137, + 0.02406391128897667, + 0.05362587049603462, + 0.0691572055220604, + 0.11089039593935013, + 0.0031801979057490826, + 0.008424827829003334, + 0.00807772483676672, + 0.0071273972280323505, + 0.0038278785068541765, + 0.006542261224240065, + 0.005002762656658888, + 0.0036595440469682217, + 0.0030259410850703716, + 0.006637997459620237, + 0.006325998809188604, + 0.008388092741370201, + 0.005469837225973606, + 0.004768135491758585, + 0.011422728188335896, + 0.0074019706808030605, + 0.008417761884629726, + 0.014338978566229343, + 0.0068265278823673725, + 0.011284666135907173, + 0.012880253605544567, + 0.010422972030937672, + 0.011381857097148895, + 0.017576957121491432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.00764006981626153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486684262752533 + ], + "opposite_logprob_tokens": [ + -1.2236684560775757 + ], + "logprob_reward": -0.3486684262752533, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486684262752533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08547432720661163, + 0.02459118328988552, + 0.06014300137758255, + 0.09397000074386597, + 0.057057712227106094, + 0.0946575328707695, + 0.4652702510356903, + 0.02082192339003086, + 0.03644382208585739, + 0.061570242047309875 + ], + "credit_map": [ + 0.03344227373600006, + 0.034340061247348785, + 0.0019889874383807182, + 0.0020696837455034256, + 0.004263099282979965, + 0.004868221934884787, + 0.002084100851789117, + 0.0024178975727409124, + 0.003940274473279715, + 0.0036357117351144552, + 0.0019768006168305874, + 0.0020303456112742424, + 0.004578865133225918, + 0.00331726367585361, + 0.0024161329492926598, + 0.0026957893278449774, + 0.006864604540169239, + 0.007149049546569586, + 0.004948471672832966, + 0.006643462460488081, + 0.01179438829421997, + 0.007071922067552805, + 0.006486632861196995, + 0.009184468537569046, + 0.012972935102880001, + 0.010393030941486359, + 0.007985186763107777, + 0.0117785120382905, + 0.012641837820410728, + 0.012423311360180378, + 0.010528543032705784, + 0.015246632508933544, + 0.007411636412143707, + 0.01223047822713852, + 0.007977652363479137, + 0.0048340666107833385, + 0.006575445644557476, + 0.007438394706696272, + 0.005356787703931332, + 0.005233251489698887, + 0.009213978424668312, + 0.019835367798805237, + 0.010887370444834232, + 0.012728751637041569, + 0.012561416253447533, + 0.011693193577229977, + 0.007076592184603214, + 0.010660870000720024, + 0.02859656885266304, + 0.037053707987070084, + 0.04493864253163338, + 0.030921563506126404, + 0.03696606680750847, + 0.06414836645126343, + 0.08679807186126709, + 0.13584724068641663, + 0.0016823380719870329, + 0.0038974029012024403, + 0.0024921020958572626, + 0.002730135340243578, + 0.002645061118528247, + 0.0036455788649618626, + 0.0020990341436117887, + 0.0016302699223160744, + 0.0029232571832835674, + 0.008071531541645527, + 0.002922459039837122, + 0.005238423123955727, + 0.0053064702078700066, + 0.004216574598103762, + 0.004067279864102602, + 0.0036978248972445726, + 0.005076099187135696, + 0.008919878862798214, + 0.0071542710065841675, + 0.014537106268107891, + 0.007147619500756264, + 0.004415753297507763, + 0.00628744438290596, + 0.008032074198126793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.3486684262752533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950118415057659 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005950118415057659, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950118415057659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1363440454006195, + 0.04096577316522598, + 0.05033553019165993, + 0.1310737133026123, + 0.11170810461044312, + 0.09686270356178284, + 0.1580706536769867, + 0.0465235598385334, + 0.0254264734685421, + 0.2026894986629486 + ], + "credit_map": [ + 0.05340476706624031, + 0.05690069496631622, + 0.002631022362038493, + 0.0026226078625768423, + 0.006528710946440697, + 0.008454274386167526, + 0.0026072002947330475, + 0.003194772405549884, + 0.0044945478439331055, + 0.00680509302765131, + 0.004061002749949694, + 0.004134954418987036, + 0.007193184457719326, + 0.005617499817162752, + 0.004075070843100548, + 0.0045844195410609245, + 0.00507213082164526, + 0.006354417651891708, + 0.0042333295568823814, + 0.006799180060625076, + 0.006680245511233807, + 0.006389590445905924, + 0.0069548990577459335, + 0.007851732894778252, + 0.016428371891379356, + 0.01365395076572895, + 0.010923927649855614, + 0.019087711349129677, + 0.016741745173931122, + 0.015625804662704468, + 0.015920981764793396, + 0.02269122190773487, + 0.013466423377394676, + 0.012907280586659908, + 0.011503186076879501, + 0.015356438234448433, + 0.011646798811852932, + 0.01564374566078186, + 0.019902056083083153, + 0.011282172985374928, + 0.010382488369941711, + 0.019690999761223793, + 0.008617042563855648, + 0.008619979955255985, + 0.01196047943085432, + 0.015995994210243225, + 0.008098859339952469, + 0.01349686086177826, + 0.01282531302422285, + 0.015634622424840927, + 0.025488808751106262, + 0.017624974250793457, + 0.012597013264894485, + 0.023382456973195076, + 0.020741507411003113, + 0.029775965958833694, + 0.0038653130177408457, + 0.008689092472195625, + 0.006128477863967419, + 0.005702713038772345, + 0.004982193931937218, + 0.008839530870318413, + 0.005342475138604641, + 0.002973764669150114, + 0.0014534619404003024, + 0.003463815664872527, + 0.0023659884463995695, + 0.0035786835942417383, + 0.0034199508372694254, + 0.004003033507615328, + 0.0044418624602258205, + 0.002699675038456917, + 0.013417492620646954, + 0.023532453924417496, + 0.01885390095412731, + 0.038746777921915054, + 0.02678833343088627, + 0.022435983642935753, + 0.02264188788831234, + 0.03627266734838486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.005950118415057659 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062971204519272 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.016062971204519272, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062971204519272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19747453927993774, + 0.03314253315329552, + 0.0779038593173027, + 0.09033950418233871, + 0.041788969188928604, + 0.03730112314224243, + 0.32677292823791504, + 0.05387558043003082, + 0.0696358010172844, + 0.07176516950130463 + ], + "credit_map": [ + 0.07894115895032883, + 0.08323575556278229, + 0.004577797371894121, + 0.00464041531085968, + 0.008475455455482006, + 0.00993893202394247, + 0.00375051680020988, + 0.003914518281817436, + 0.004626154433935881, + 0.0059769246727228165, + 0.0030530488584190607, + 0.003868234110996127, + 0.003584859659895301, + 0.0040381536819040775, + 0.002613408723846078, + 0.005381749477237463, + 0.007856952957808971, + 0.0087008411064744, + 0.006781446747481823, + 0.009487533941864967, + 0.013512208126485348, + 0.008646423928439617, + 0.008782424032688141, + 0.01413603313267231, + 0.011434280313551426, + 0.010301622562110424, + 0.007985075004398823, + 0.013187626376748085, + 0.012439657934010029, + 0.010895670391619205, + 0.009960466995835304, + 0.014135103672742844, + 0.004153928719460964, + 0.006782252341508865, + 0.004157881252467632, + 0.004926170222461224, + 0.004071604926139116, + 0.004443309735506773, + 0.007507198955863714, + 0.005746626295149326, + 0.0037466150242835283, + 0.0073701911605894566, + 0.005011030938476324, + 0.005704694893211126, + 0.005410275887697935, + 0.005045068450272083, + 0.002644096966832876, + 0.002369149588048458, + 0.025639431551098824, + 0.03136560693383217, + 0.0571620799601078, + 0.04054224118590355, + 0.027910424396395683, + 0.04599114507436752, + 0.04696796461939812, + 0.05119403824210167, + 0.004000748973339796, + 0.0064633372239768505, + 0.011607197113335133, + 0.008697524666786194, + 0.00499786576256156, + 0.0067826672457158566, + 0.006609695963561535, + 0.004716543946415186, + 0.003799884347245097, + 0.005230980925261974, + 0.006640393286943436, + 0.007805967703461647, + 0.010737408883869648, + 0.014071343466639519, + 0.01435281615704298, + 0.006997013930231333, + 0.00789481308311224, + 0.013420338742434978, + 0.007631207816302776, + 0.012193506583571434, + 0.009306458756327629, + 0.0059656910598278046, + 0.006864322815090418, + 0.008488819934427738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.016062971204519272 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This medium shot, captured from a slightly low angle with a handheld camera that subtly sways and undulates in sync with the water's movement, focuses on an elderly Asian man with thinning, silver-gray hair and a deeply wrinkled face. He sits in a small, weathered wooden boat, deliberately lowering his center of gravity and bending his knees to counteract a sudden numbness. His rough, gnarled hands, with prominent veins and calloused skin, instinctively grasp the gunwale of the dark, damp boat, his knuckles pale from the effort. The boat lurches on the unstable, dark waters, and the turbulent waves crash against the aged wood. The scene is illuminated by the dim, unfocused light of an overcast sky, creating a cool, desaturated palette of deep blues, grays, and dark browns that echo the damp boat. The man's expression is one of intense, quiet concentration, not fear, his gaze fixed on the turbulent waters ahead, creating a tense and somber atmosphere of profound and enduring resilience against the elements.", + "seed": 20781, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_12/index_4427_seed_20781_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": -0.022211696952581406, + "vlm_logprob": -0.022211696952581406, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.022211696952581406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01827123388648033 + ], + "opposite_logprob_tokens": [ + -4.143271446228027 + ], + "logprob_reward": -0.01827123388648033, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01827123388648033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10158750414848328, + 0.0636587142944336, + 0.08407265692949295, + 0.08227376639842987, + 0.06519997864961624, + 0.05448600649833679, + 0.10719384998083115, + 0.15863846242427826, + 0.20146845281124115, + 0.08142054080963135 + ], + "credit_map": [ + 0.006799605209380388, + 0.01322297751903534, + 0.015730682760477066, + 0.015162433497607708, + 0.016883613541722298, + 0.007476978003978729, + 0.010637949220836163, + 0.01567326858639717, + 0.005147538147866726, + 0.007370424922555685, + 0.006755057256668806, + 0.009149125777184963, + 0.009182965382933617, + 0.006370024289935827, + 0.010335865430533886, + 0.009347716346383095, + 0.0076374271884560585, + 0.007331945467740297, + 0.010556930676102638, + 0.00915700662881136, + 0.00683238822966814, + 0.00469902204349637, + 0.022447414696216583, + 0.015410522930324078, + 0.007585415616631508, + 0.011331162415444851, + 0.008394205942749977, + 0.01160674449056387, + 0.012823169119656086, + 0.008150685578584671, + 0.010926970280706882, + 0.011455418542027473, + 0.006650575902312994, + 0.011866752989590168, + 0.006191604770720005, + 0.003979354631155729, + 0.008465015329420567, + 0.0053171212784945965, + 0.009857650846242905, + 0.012871905229985714, + 0.00689715938642621, + 0.011203888803720474, + 0.007218451704829931, + 0.005835851654410362, + 0.009774722158908844, + 0.004798292648047209, + 0.00499351741746068, + 0.003764124820008874, + 0.01104185450822115, + 0.022281475365161896, + 0.00581259373575449, + 0.01004638522863388, + 0.023637808859348297, + 0.016306236386299133, + 0.007243188563734293, + 0.010824312455952168, + 0.013205583207309246, + 0.022342396900057793, + 0.01750011183321476, + 0.019748756662011147, + 0.04239128902554512, + 0.018889591097831726, + 0.01580119878053665, + 0.008759557269513607, + 0.048102058470249176, + 0.02937646210193634, + 0.031614527106285095, + 0.014503847807645798, + 0.023196054622530937, + 0.019639689475297928, + 0.009818094782531261, + 0.025217726826667786, + 0.007294840645045042, + 0.010066685266792774, + 0.007486932910978794, + 0.009381930343806744, + 0.006653597578406334, + 0.011412268504500389, + 0.016974899917840958, + 0.012149399146437645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01827123388648033 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046497792936861515 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.0046497792936861515, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046497792936861515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29145997762680054, + 0.20062987506389618, + 0.08062203228473663, + 0.06866590678691864, + 0.023021897301077843, + 0.12918969988822937, + 0.023171981796622276, + 0.09465766698122025, + 0.03351154923439026, + 0.055069450289011 + ], + "credit_map": [ + 0.02379807084798813, + 0.041669394820928574, + 0.02495766244828701, + 0.029490508139133453, + 0.030254624783992767, + 0.0524265356361866, + 0.01925203949213028, + 0.06961110979318619, + 0.01748267374932766, + 0.02139618620276451, + 0.023623071610927582, + 0.027000872418284416, + 0.0239397119730711, + 0.020448459312319756, + 0.03707745298743248, + 0.029661424458026886, + 0.007715081330388784, + 0.006975248456001282, + 0.008314517326653004, + 0.010983245447278023, + 0.007039146963506937, + 0.004566935822367668, + 0.018740903586149216, + 0.01628694497048855, + 0.005383639130741358, + 0.008701867423951626, + 0.00883556343615055, + 0.014928608201444149, + 0.00845665205270052, + 0.004316031001508236, + 0.008860070258378983, + 0.00918346457183361, + 0.002084752544760704, + 0.004814154468476772, + 0.0030053462833166122, + 0.0018225269159302115, + 0.00340967602096498, + 0.0019390956731513143, + 0.003568730317056179, + 0.0023776113521307707, + 0.018785379827022552, + 0.016522908583283424, + 0.014769193716347218, + 0.013906383886933327, + 0.017337588593363762, + 0.015191597864031792, + 0.013917728327214718, + 0.01875890977680683, + 0.0022158008068799973, + 0.0055056046694517136, + 0.0018276114715263247, + 0.0020254889968782663, + 0.0036373171024024487, + 0.003926358185708523, + 0.0012931235833093524, + 0.0027406758163124323, + 0.008317459374666214, + 0.013460769318044186, + 0.008811011910438538, + 0.011597618460655212, + 0.019695565104484558, + 0.012021386064589024, + 0.012204200960695744, + 0.008549644611775875, + 0.005464071873575449, + 0.004462650045752525, + 0.003502587554976344, + 0.002830480458214879, + 0.004510829690843821, + 0.0038269672077149153, + 0.005701199173927307, + 0.0032127618324011564, + 0.005443935748189688, + 0.005591091699898243, + 0.004146493971347809, + 0.006525027099996805, + 0.0042028771713376045, + 0.008506321348249912, + 0.012015825137495995, + 0.008637872524559498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0046497792936861515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062039099633693695 + ], + "opposite_logprob_tokens": [ + -2.8120391368865967 + ], + "logprob_reward": -0.062039099633693695, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062039099633693695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12589599192142487, + 0.05375754460692406, + 0.09434787929058075, + 0.07571186125278473, + 0.027182871475815773, + 0.08492061495780945, + 0.04443714767694473, + 0.3136350214481354, + 0.09406032413244247, + 0.08605071157217026 + ], + "credit_map": [ + 0.010948621667921543, + 0.01576792262494564, + 0.019099202007055283, + 0.015488428995013237, + 0.015209881588816643, + 0.01978956162929535, + 0.010598848573863506, + 0.018993526697158813, + 0.004044830333441496, + 0.005216961260885, + 0.005840257741510868, + 0.00827010441571474, + 0.006818635389208794, + 0.004715126473456621, + 0.010675895027816296, + 0.00817573070526123, + 0.0051030307076871395, + 0.007366719655692577, + 0.008495553396642208, + 0.01175964530557394, + 0.007246430963277817, + 0.007850535213947296, + 0.026487234979867935, + 0.020038731396198273, + 0.012774594128131866, + 0.010944067500531673, + 0.008603859692811966, + 0.009451976045966148, + 0.008965695276856422, + 0.007639564573764801, + 0.00791835505515337, + 0.009413751773536205, + 0.002358050085604191, + 0.004156428389251232, + 0.003558861091732979, + 0.0030011809431016445, + 0.004154528491199017, + 0.0019241933478042483, + 0.005183625500649214, + 0.0028460046742111444, + 0.008375915698707104, + 0.013802778907120228, + 0.009922935627400875, + 0.0122126629576087, + 0.014870726503431797, + 0.009203062392771244, + 0.010116009041666985, + 0.0064165289513766766, + 0.005004985257983208, + 0.008734706789255142, + 0.004164042416960001, + 0.004191823769360781, + 0.0084941815584898, + 0.006014338694512844, + 0.0027962245512753725, + 0.005036843940615654, + 0.02667403034865856, + 0.03632329776883125, + 0.018708063289523125, + 0.029264716431498528, + 0.03179493546485901, + 0.07771816104650497, + 0.028056414797902107, + 0.06509540975093842, + 0.020066028460860252, + 0.010571373626589775, + 0.009524524211883545, + 0.006695105694234371, + 0.008449329994618893, + 0.010486645624041557, + 0.007795657962560654, + 0.02047165296971798, + 0.00752452714368701, + 0.010221521370112896, + 0.011259308084845543, + 0.012258917093276978, + 0.00552252447232604, + 0.00878049060702324, + 0.01833222433924675, + 0.012151196599006653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.062039099633693695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.190488800406456, + 0.08157047629356384, + 0.17252151668071747, + 0.03726227954030037, + 0.0240758266299963, + 0.1883317530155182, + 0.023958435282111168, + 0.12551112473011017, + 0.11366788297891617, + 0.04261188581585884 + ], + "credit_map": [ + 0.01270744577050209, + 0.02460041455924511, + 0.020898867398500443, + 0.023847416043281555, + 0.02511000633239746, + 0.02916690520942211, + 0.015319650061428547, + 0.03883810713887215, + 0.005382767878472805, + 0.007110493257641792, + 0.007033910136669874, + 0.011995404958724976, + 0.010122964158654213, + 0.008753029629588127, + 0.019275149330496788, + 0.011896762996912003, + 0.011539204977452755, + 0.014596079476177692, + 0.027104221284389496, + 0.026834337040781975, + 0.01622353494167328, + 0.01202748529613018, + 0.030717337504029274, + 0.03347932547330856, + 0.0031368073541671038, + 0.005706800613552332, + 0.004256002604961395, + 0.007545209489762783, + 0.0038562931586056948, + 0.002777714282274246, + 0.00519905798137188, + 0.004784393589943647, + 0.0022713406942784786, + 0.0036137434653937817, + 0.002403967548161745, + 0.0020100122783333063, + 0.004814879037439823, + 0.0023453838657587767, + 0.0037496741861104965, + 0.0028668285813182592, + 0.019505662843585014, + 0.027671977877616882, + 0.021828727796673775, + 0.025754256173968315, + 0.031209688633680344, + 0.020595552399754524, + 0.02352675050497055, + 0.018239151686429977, + 0.002622675383463502, + 0.0030052433721721172, + 0.002278271596878767, + 0.001691367942839861, + 0.005324619356542826, + 0.004063715226948261, + 0.0028469893150031567, + 0.0021255549509078264, + 0.009552757255733013, + 0.013000274077057838, + 0.011912142857909203, + 0.015528897754848003, + 0.028115808963775635, + 0.018543081358075142, + 0.014141291379928589, + 0.014716888777911663, + 0.029596226289868355, + 0.013462691567838192, + 0.015499611385166645, + 0.006756929215043783, + 0.013579828664660454, + 0.01294146291911602, + 0.006114382296800613, + 0.015716763213276863, + 0.004502943716943264, + 0.004481158684939146, + 0.003276569303125143, + 0.003907914273440838, + 0.003967905417084694, + 0.008344847708940506, + 0.0082008121535182, + 0.005929735954850912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.002819731365889311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Asian man with thinning silver-gray hair and a deeply wrinkled face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man sitting in a small, weathered wooden boat with rough, gnarled hands grasping the gunwale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man actively lowering his center of gravity and bending his knees to counteract instability?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by dark, damp, turbulent water with waves crashing against the boat under dim, overcast lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot captured from a slightly low angle with a handheld sway and undulation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327864058315754 + ], + "opposite_logprob_tokens": [ + -3.7732787132263184 + ], + "logprob_reward": -0.02327864058315754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327864058315754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20719894766807556, + 0.1513696312904358, + 0.05505668371915817, + 0.10958795994520187, + 0.03837328776717186, + 0.04147983342409134, + 0.05185168609023094, + 0.17400674521923065, + 0.06652611494064331, + 0.10454920679330826 + ], + "credit_map": [ + 0.015297922305762768, + 0.025258762761950493, + 0.021460626274347305, + 0.019369404762983322, + 0.019279783591628075, + 0.038639381527900696, + 0.01237365510314703, + 0.05551941692829132, + 0.011641886085271835, + 0.009661165066063404, + 0.015777653083205223, + 0.028853889554739, + 0.018154263496398926, + 0.008796902373433113, + 0.03710756078362465, + 0.02137630060315132, + 0.004840725101530552, + 0.004582224879413843, + 0.008280857466161251, + 0.008512396365404129, + 0.0037079520989209414, + 0.004218139685690403, + 0.010869413614273071, + 0.010044979862868786, + 0.00609215022996068, + 0.01313296239823103, + 0.0116495992988348, + 0.02585890144109726, + 0.010426138527691364, + 0.010489488951861858, + 0.018012909218668938, + 0.013925809413194656, + 0.0027684103697538376, + 0.007255571894347668, + 0.003917493391782045, + 0.002766277175396681, + 0.004588300362229347, + 0.004568828735500574, + 0.007981889881193638, + 0.0045265136286616325, + 0.005948585458099842, + 0.007417233660817146, + 0.004365506116300821, + 0.005311466287821531, + 0.005484383553266525, + 0.0035687729250639677, + 0.003787148743867874, + 0.005596736446022987, + 0.005249855108559132, + 0.010980550199747086, + 0.004214073531329632, + 0.0035904766991734505, + 0.008294482715427876, + 0.00881284661591053, + 0.005190586671233177, + 0.005518813617527485, + 0.012554184533655643, + 0.022765882313251495, + 0.01675226353108883, + 0.01989828795194626, + 0.040145423263311386, + 0.023625416681170464, + 0.01911926083266735, + 0.019146014004945755, + 0.014000225812196732, + 0.00648054014891386, + 0.009111468680202961, + 0.0038993016351014376, + 0.007033905480057001, + 0.007992628030478954, + 0.004735354799777269, + 0.013272684067487717, + 0.010806042701005936, + 0.013757545500993729, + 0.008093633688986301, + 0.00980292446911335, + 0.009467159397900105, + 0.017200781032443047, + 0.020283658057451248, + 0.015137463808059692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02327864058315754 + } + ], + "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_13/index_tensor([6012, 6012])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_tensor([6012, 6012])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ab52273a395fe73010d397519bed92d71c6d33c2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_tensor([6012, 6012])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle shot captures the scene, rendered in the friendly, colorful style of Pixar animation, focusing on the perfect centerpiece of a lush, mossy forest floor. The scene is bathed in warm, dappled sunlight filtering through the unseen treetops, creating softly edged shadows. A small mound amidst the emerald green moss begins to pulsate, and then a captivating fungus powerfully thrusts into view. Its creamy white stem, dotted with cheerful reddish-brown spots, rapidly elongates upward. The cap, initially a compact, vibrant pumpkin-orange sphere, gracefully unfolds in time-lapse photography, its smooth, saturated surface expanding outward to reveal its full, classic mushroom form. As it reaches its final, proud stance, the entire mushroom exhibits a subtle, endearing sway, blending into the magical forest scene with a palpable sense of life and whimsy.", + "seed": 111, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_6012_seed_111_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": -0.0033726199413649737, + "vlm_logprob": -0.0033726199413649737, + "vlm_yes_no_margin": 6.2, + "vlm_token_credit": -0.0033726199413649737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0053331246599555016 + ], + "opposite_logprob_tokens": [ + -6.380332946777344 + ], + "logprob_reward": -0.0053331246599555016, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0053331246599555016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1939605474472046, + 0.09804675728082657, + 0.05503325164318085, + 0.10480824112892151, + 0.044761065393686295, + 0.09413652122020721, + 0.07126449793577194, + 0.10869914293289185, + 0.05747271701693535, + 0.17181725800037384 + ], + "credit_map": [ + 0.019017092883586884, + 0.03650972247123718, + 0.01947990618646145, + 0.015232718549668789, + 0.02764575555920601, + 0.03456640616059303, + 0.024269983172416687, + 0.017238978296518326, + 0.01053523737937212, + 0.012391440570354462, + 0.010214243084192276, + 0.010289918631315231, + 0.016761623322963715, + 0.014873694628477097, + 0.013357134535908699, + 0.009623462334275246, + 0.007383222691714764, + 0.008871187455952168, + 0.004200650844722986, + 0.004716791212558746, + 0.011822526343166828, + 0.008047209121286869, + 0.005296696443110704, + 0.004694966599345207, + 0.012495793402194977, + 0.016432391479611397, + 0.009009265340864658, + 0.008667404763400555, + 0.025723231956362724, + 0.0149689931422472, + 0.008075579069554806, + 0.009435581974685192, + 0.007191543933004141, + 0.005993944592773914, + 0.004078371915966272, + 0.004687524866312742, + 0.008706540800631046, + 0.005830385722219944, + 0.0042346324771642685, + 0.004038121085613966, + 0.01511929091066122, + 0.009557828307151794, + 0.007096054032444954, + 0.008401371538639069, + 0.014789056032896042, + 0.014964842237532139, + 0.015039147809147835, + 0.009168935939669609, + 0.010304837487637997, + 0.010295243933796883, + 0.007822833955287933, + 0.005526955705136061, + 0.009402500465512276, + 0.01088644564151764, + 0.006643473636358976, + 0.010382206179201603, + 0.01857578568160534, + 0.007371796295046806, + 0.008930456824600697, + 0.006657890509814024, + 0.010471814312040806, + 0.03962957113981247, + 0.006307859439402819, + 0.010753961279988289, + 0.004960883874446154, + 0.0037573559675365686, + 0.010091443546116352, + 0.004333145450800657, + 0.004373636562377214, + 0.02059091255068779, + 0.003973854705691338, + 0.005391484126448631, + 0.02756160870194435, + 0.01172129437327385, + 0.057796865701675415, + 0.01681852713227272, + 0.009918984957039356, + 0.024543223902583122, + 0.007277186494320631, + 0.016179563477635384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0053331246599555016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009260656661354005 + ], + "opposite_logprob_tokens": [ + -7.0009260177612305 + ], + "logprob_reward": -0.0009260656661354005, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009260656661354005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19879576563835144, + 0.1549433320760727, + 0.06328736990690231, + 0.07549493759870529, + 0.03533707186579704, + 0.029263239353895187, + 0.07719261199235916, + 0.047831159085035324, + 0.18020476400852203, + 0.1376497596502304 + ], + "credit_map": [ + 0.01737120747566223, + 0.023228265345096588, + 0.022731656208634377, + 0.019928699359297752, + 0.025882815942168236, + 0.028475841507315636, + 0.02051723375916481, + 0.04066005349159241, + 0.02210589125752449, + 0.017268341034650803, + 0.01382004376500845, + 0.017386341467499733, + 0.028511490672826767, + 0.021684061735868454, + 0.012525559403002262, + 0.02164161391556263, + 0.010399789549410343, + 0.007875937968492508, + 0.006440074648708105, + 0.005173008423298597, + 0.012891313061118126, + 0.009256967343389988, + 0.005239309277385473, + 0.006010967772454023, + 0.0074135069735348225, + 0.013180240988731384, + 0.004041274078190327, + 0.0054226587526500225, + 0.022677797824144363, + 0.010311541147530079, + 0.006790808402001858, + 0.005657109897583723, + 0.006296534091234207, + 0.004121166188269854, + 0.0030514791142195463, + 0.003530940506607294, + 0.007637656293809414, + 0.004564779810607433, + 0.002714486327022314, + 0.0034200299996882677, + 0.00456391042098403, + 0.0031640371307730675, + 0.0028387608472257853, + 0.002671396592631936, + 0.0043288725428283215, + 0.005409889854490757, + 0.0024766044225543737, + 0.003809768008068204, + 0.01314238365739584, + 0.007193928584456444, + 0.005766520742326975, + 0.005103694275021553, + 0.009455093182623386, + 0.019096139818429947, + 0.008376889862120152, + 0.009057961404323578, + 0.00822222325950861, + 0.004194972570985556, + 0.006366935092955828, + 0.004514956846833229, + 0.004898271523416042, + 0.008483023382723331, + 0.0060458495281636715, + 0.005104925483465195, + 0.017061227932572365, + 0.012669609859585762, + 0.029978180304169655, + 0.013361052609980106, + 0.012850132770836353, + 0.05857033655047417, + 0.01924930326640606, + 0.016464920714497566, + 0.026062658056616783, + 0.011588387191295624, + 0.03854916989803314, + 0.012874806299805641, + 0.01260724663734436, + 0.018656382337212563, + 0.005543454084545374, + 0.011767652817070484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0009260656661354005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010448002722114325 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010448002722114325, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010448002722114325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3488238453865051, + 0.07178109139204025, + 0.2081274539232254, + 0.06366794556379318, + 0.047772061079740524, + 0.0365385040640831, + 0.04695095494389534, + 0.031392015516757965, + 0.023466669023036957, + 0.12147938460111618 + ], + "credit_map": [ + 0.024893468245863914, + 0.060381218791007996, + 0.018207190558314323, + 0.023226751014590263, + 0.08695106953382492, + 0.058264125138521194, + 0.05301622673869133, + 0.023883799090981483, + 0.0061093587428331375, + 0.007438137196004391, + 0.004422878380864859, + 0.006868482101708651, + 0.0183509960770607, + 0.00894220732152462, + 0.012774214148521423, + 0.006874811369925737, + 0.025948721915483475, + 0.030240710824728012, + 0.015255948528647423, + 0.01965269073843956, + 0.0423644594848156, + 0.03026820532977581, + 0.023113803938031197, + 0.021282926201820374, + 0.007175459060817957, + 0.010213528759777546, + 0.004661067388951778, + 0.0054792058654129505, + 0.016051439568400383, + 0.008662124164402485, + 0.006789316423237324, + 0.004635802935808897, + 0.005477820057421923, + 0.006161772180348635, + 0.003132112789899111, + 0.003422355279326439, + 0.0126493526622653, + 0.0077756973914802074, + 0.005747360642999411, + 0.003405593102797866, + 0.005633848253637552, + 0.005196882877498865, + 0.002703831298276782, + 0.0036410221364349127, + 0.0070687467232346535, + 0.005052397958934307, + 0.0034328019246459007, + 0.0038089719600975513, + 0.006006936077028513, + 0.00758232781663537, + 0.004532027989625931, + 0.0026616910472512245, + 0.011424495838582516, + 0.005809738766402006, + 0.0053102243691682816, + 0.0036235146690160036, + 0.004973637405782938, + 0.0019215097418054938, + 0.0029166729655116796, + 0.0020235739648342133, + 0.002831714227795601, + 0.011227628216147423, + 0.002693929011002183, + 0.0028033489361405373, + 0.0027688564732670784, + 0.002915971679612994, + 0.002101396443322301, + 0.002311522141098976, + 0.0025055634323507547, + 0.0036239554174244404, + 0.004721374250948429, + 0.0025180275551974773, + 0.012456124648451805, + 0.013370529748499393, + 0.017337238416075706, + 0.00897736195474863, + 0.02516951411962509, + 0.01763628050684929, + 0.013764610514044762, + 0.012767736800014973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0010448002722114325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1783008575439453, + 0.1754491776227951, + 0.25975722074508667, + 0.03892923891544342, + 0.031203078106045723, + 0.028913116082549095, + 0.032423749566078186, + 0.017452547326683998, + 0.014291820116341114, + 0.2232791781425476 + ], + "credit_map": [ + 0.011345994658768177, + 0.026215653866529465, + 0.009530901908874512, + 0.012817418202757835, + 0.04791964590549469, + 0.02460142970085144, + 0.03388100489974022, + 0.011988814920186996, + 0.01649572141468525, + 0.01971537061035633, + 0.008463156409561634, + 0.013014177791774273, + 0.05562916025519371, + 0.027026036754250526, + 0.021946052089333534, + 0.013159506022930145, + 0.03182428330183029, + 0.03843825310468674, + 0.016604475677013397, + 0.028120703995227814, + 0.04914636164903641, + 0.0376291498541832, + 0.03161012753844261, + 0.02638385258615017, + 0.005071135237812996, + 0.005603513214737177, + 0.0037036908324807882, + 0.0035332688130438328, + 0.008720247074961662, + 0.005463979206979275, + 0.0039010930340737104, + 0.0029323138296604156, + 0.004011439625173807, + 0.0038172765634953976, + 0.002240539062768221, + 0.00269553204998374, + 0.007948214188218117, + 0.003816136159002781, + 0.0035270382650196552, + 0.003146903356537223, + 0.005853657145053148, + 0.003352473024278879, + 0.002365810563787818, + 0.0027587253134697676, + 0.004014914855360985, + 0.0048518311232328415, + 0.0028270799666643143, + 0.0028886229265481234, + 0.005463641602545977, + 0.0027386494912207127, + 0.003245435655117035, + 0.002083896892145276, + 0.003489298280328512, + 0.00979460310190916, + 0.002155146561563015, + 0.003453077981248498, + 0.002175539033487439, + 0.001953011378645897, + 0.001269439933821559, + 0.0018232724396511912, + 0.0020529399625957012, + 0.0020482924301177263, + 0.00486254645511508, + 0.0012675070902332664, + 0.0016503462102264166, + 0.0020621842704713345, + 0.0012912781676277518, + 0.0015454075764864683, + 0.0018993515986949205, + 0.0018929974175989628, + 0.002492758212611079, + 0.0014574979431927204, + 0.03164764866232872, + 0.03087799809873104, + 0.0497148223221302, + 0.019932879135012627, + 0.030236972495913506, + 0.02508413791656494, + 0.017231885343790054, + 0.018552830442786217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.00361420470289886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944904405623674 + ], + "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": [ + 0.1284845918416977, + 0.13027986884117126, + 0.1260729730129242, + 0.06977459043264389, + 0.04169858247041702, + 0.08936901390552521, + 0.09216447174549103, + 0.1156877875328064, + 0.09892279654741287, + 0.10754526406526566 + ], + "credit_map": [ + 0.010108358226716518, + 0.016480473801493645, + 0.007571541704237461, + 0.009985753335058689, + 0.03487502038478851, + 0.02002459578216076, + 0.017336931079626083, + 0.012101927772164345, + 0.022274941205978394, + 0.019190356135368347, + 0.012887522578239441, + 0.013207835145294666, + 0.01967548206448555, + 0.01936446502804756, + 0.00979604758322239, + 0.013883234933018684, + 0.015384211204946041, + 0.01916666515171528, + 0.009295162744820118, + 0.01151701807975769, + 0.026020782068371773, + 0.018505392596125603, + 0.013579824939370155, + 0.012603923678398132, + 0.005936551373451948, + 0.012701848521828651, + 0.004164885729551315, + 0.004732076544314623, + 0.021411899477243423, + 0.009103868156671524, + 0.006680735852569342, + 0.005042729899287224, + 0.007462240289896727, + 0.003963380586355925, + 0.00493852561339736, + 0.004104321822524071, + 0.005070494953542948, + 0.007761269807815552, + 0.003494192846119404, + 0.004904160741716623, + 0.019856125116348267, + 0.005851335823535919, + 0.010595140047371387, + 0.009879241697490215, + 0.006778283976018429, + 0.016287943348288536, + 0.0070799547247588634, + 0.013040995225310326, + 0.017001336440443993, + 0.004998310469090939, + 0.009136681444942951, + 0.005607410799711943, + 0.004831100348383188, + 0.03710860759019852, + 0.005601774901151657, + 0.007879259996116161, + 0.016245583072304726, + 0.006820331327617168, + 0.01111946627497673, + 0.009397857822477818, + 0.00703069195151329, + 0.03857174888253212, + 0.017608966678380966, + 0.008893142454326153, + 0.011015512980520725, + 0.011086690239608288, + 0.018609408289194107, + 0.008909881114959717, + 0.010079724714159966, + 0.01678837649524212, + 0.014230696484446526, + 0.008202515542507172, + 0.010798945091664791, + 0.01470863539725542, + 0.010662664659321308, + 0.007702369708567858, + 0.022091615945100784, + 0.023203538730740547, + 0.00866683479398489, + 0.009710663929581642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.005944904405623674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures the scene, rendered in the friendly, colorful style of Pixar animation, focusing on the perfect centerpiece of a lush, mossy forest floor. The scene is bathed in warm, dappled sunlight filtering through the unseen treetops, creating softly edged shadows. A small mound amidst the emerald green moss begins to pulsate, and then a captivating fungus powerfully thrusts into view. Its creamy white stem, dotted with cheerful reddish-brown spots, rapidly elongates upward. The cap, initially a compact, vibrant pumpkin-orange sphere, gracefully unfolds in time-lapse photography, its smooth, saturated surface expanding outward to reveal its full, classic mushroom form. As it reaches its final, proud stance, the entire mushroom exhibits a subtle, endearing sway, blending into the magical forest scene with a palpable sense of life and whimsy.", + "seed": 112, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_6012_seed_112_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": -0.0037825133884325625, + "vlm_logprob": -0.0037825133884325625, + "vlm_yes_no_margin": 6.1, + "vlm_token_credit": -0.0037825133884325625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006772303022444248 + ], + "opposite_logprob_tokens": [ + -6.506772518157959 + ], + "logprob_reward": -0.006772303022444248, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.006772303022444248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10176686942577362, + 0.03241552785038948, + 0.0069569372572004795, + 0.01674293540418148, + 0.006912799086421728, + 0.3252497613430023, + 0.018103603273630142, + 0.2556215524673462, + 0.16320428252220154, + 0.07302574813365936 + ], + "credit_map": [ + 0.010317627340555191, + 0.013476461172103882, + 0.0052948747761547565, + 0.006635836325585842, + 0.01662028208374977, + 0.027596332132816315, + 0.008139299228787422, + 0.013686145655810833, + 0.003907744772732258, + 0.0022422047331929207, + 0.002476761117577553, + 0.0016726916655898094, + 0.0021332085598260164, + 0.011948571540415287, + 0.005420591216534376, + 0.0026137514505535364, + 0.0009232679149135947, + 0.0013232307974249125, + 0.000698705785907805, + 0.0007042500656098127, + 0.001033857581205666, + 0.0011177131673321128, + 0.0005264076753519475, + 0.000629504444077611, + 0.0017149863997474313, + 0.0018770399037748575, + 0.0007903584628365934, + 0.0016567237908020616, + 0.0017634376417845488, + 0.005426293238997459, + 0.0025961967185139656, + 0.0009178974432870746, + 0.0008145756437443197, + 0.0010223897406831384, + 0.000648361281491816, + 0.0007706318865530193, + 0.001383851165883243, + 0.0009575483854860067, + 0.0007259061676450074, + 0.0005895341746509075, + 0.03342040628194809, + 0.06272120773792267, + 0.023476319387555122, + 0.026385657489299774, + 0.06988970935344696, + 0.061336446553468704, + 0.027258966118097305, + 0.02076103165745735, + 0.002385047497227788, + 0.002475235378369689, + 0.0017214781837537885, + 0.0015175306471064687, + 0.004271973390132189, + 0.002218160079792142, + 0.0016417723381891847, + 0.001872402848675847, + 0.037029165774583817, + 0.04812826216220856, + 0.015522375702857971, + 0.025944529101252556, + 0.06128988415002823, + 0.029719093814492226, + 0.021604331210255623, + 0.01638389565050602, + 0.02258087322115898, + 0.03360116481781006, + 0.009806769900023937, + 0.01581554487347603, + 0.03732932358980179, + 0.018635425716638565, + 0.012751592323184013, + 0.012683589942753315, + 0.010078974068164825, + 0.014532911591231823, + 0.004010790027678013, + 0.006776620168238878, + 0.018276061862707138, + 0.008817818947136402, + 0.006034485995769501, + 0.004498082213103771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.006772303022444248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019480077316984534 + ], + "opposite_logprob_tokens": [ + -6.25194787979126 + ], + "logprob_reward": -0.0019480077316984534, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019480077316984534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11041277647018433, + 0.10790085047483444, + 0.019681857898831367, + 0.0347893051803112, + 0.02242681384086609, + 0.11240356415510178, + 0.0675431415438652, + 0.273285448551178, + 0.1289060264825821, + 0.12265018373727798 + ], + "credit_map": [ + 0.009396670386195183, + 0.016269976273179054, + 0.006597294472157955, + 0.010638735257089138, + 0.018925132229924202, + 0.021320508792996407, + 0.013069002889096737, + 0.014195451512932777, + 0.01286275964230299, + 0.006357653997838497, + 0.005122927017509937, + 0.0038117249496281147, + 0.007971894927322865, + 0.044266797602176666, + 0.020395321771502495, + 0.007111777551472187, + 0.0022723921574652195, + 0.0027855101507157087, + 0.0029769521206617355, + 0.0020819997880607843, + 0.002475355751812458, + 0.002496374072507024, + 0.002532700775191188, + 0.0020605716854333878, + 0.004455917980521917, + 0.00377930817194283, + 0.00472791725769639, + 0.0036281864158809185, + 0.004516775719821453, + 0.005987189710140228, + 0.004093483090400696, + 0.003600527998059988, + 0.00195717578753829, + 0.0028605542611330748, + 0.0029919417575001717, + 0.0026008605491369963, + 0.004042536485940218, + 0.0035283686593174934, + 0.002138579962775111, + 0.0023067963775247335, + 0.01122212503105402, + 0.02273288555443287, + 0.008406823500990868, + 0.011085718870162964, + 0.028262227773666382, + 0.01404593326151371, + 0.008565673604607582, + 0.008082171902060509, + 0.007126686163246632, + 0.011683606542646885, + 0.005767607595771551, + 0.008559172973036766, + 0.013206598348915577, + 0.007648970931768417, + 0.008568189106881618, + 0.004982306156307459, + 0.040574248880147934, + 0.05032997205853462, + 0.017789136618375778, + 0.024913320317864418, + 0.06264293938875198, + 0.03541553020477295, + 0.021384520456194878, + 0.02023577317595482, + 0.017195597290992737, + 0.023288147523999214, + 0.00885001290589571, + 0.014872546307742596, + 0.02809765934944153, + 0.014271693304181099, + 0.013106608763337135, + 0.00922375451773405, + 0.017365824431180954, + 0.02560807205736637, + 0.006407052278518677, + 0.011064311489462852, + 0.03069506213068962, + 0.014276725240051746, + 0.010574397630989552, + 0.006658738013356924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0019480077316984534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015159078175202012 + ], + "opposite_logprob_tokens": [ + -6.501515865325928 + ], + "logprob_reward": -0.0015159078175202012, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015159078175202012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04962268844246864, + 0.03708461672067642, + 0.05069401487708092, + 0.026053648442029953, + 0.03049314208328724, + 0.31509387493133545, + 0.05499005317687988, + 0.3683677017688751, + 0.05237361416220665, + 0.015226621180772781 + ], + "credit_map": [ + 0.005681470036506653, + 0.010502641089260578, + 0.00415312172845006, + 0.006049481686204672, + 0.00713084964081645, + 0.008084124885499477, + 0.004279963206499815, + 0.0037410398945212364, + 0.0033820588141679764, + 0.002825656672939658, + 0.002419714815914631, + 0.0020590887870639563, + 0.0030822157859802246, + 0.01231566071510315, + 0.008373717777431011, + 0.0026265066117048264, + 0.004373890347778797, + 0.0045670559629797935, + 0.003029649844393134, + 0.0036402088589966297, + 0.004102801904082298, + 0.018625160679221153, + 0.00919825118035078, + 0.003157002152875066, + 0.003153035417199135, + 0.005042853765189648, + 0.0026711460668593645, + 0.0031113275326788425, + 0.004648948088288307, + 0.0029297482687979937, + 0.0026006014086306095, + 0.001895990688353777, + 0.0029759746976196766, + 0.003759579034522176, + 0.001757077407091856, + 0.0027271911967545748, + 0.0047633955255150795, + 0.008499759249389172, + 0.004029419273138046, + 0.001980749424546957, + 0.031012963503599167, + 0.06106359884142876, + 0.022280480712652206, + 0.02693062461912632, + 0.07198963314294815, + 0.055187880992889404, + 0.025998001918196678, + 0.0206307265907526, + 0.006200629752129316, + 0.007480853237211704, + 0.0050548892468214035, + 0.007188587449491024, + 0.008868592791259289, + 0.010363460518419743, + 0.005807213950902224, + 0.004025834612548351, + 0.05374453589320183, + 0.0691886693239212, + 0.02258249744772911, + 0.036036863923072815, + 0.0865505263209343, + 0.044723015278577805, + 0.030153535306453705, + 0.025388065725564957, + 0.006493567954748869, + 0.0071390108205378056, + 0.00489593343809247, + 0.004820867441594601, + 0.01009448990225792, + 0.007783190347254276, + 0.005902326200157404, + 0.005244235973805189, + 0.0019949115812778473, + 0.00274535920470953, + 0.0010850938269868493, + 0.0014591597719118, + 0.0037065341603010893, + 0.001730938907712698, + 0.0015318061923608184, + 0.0009728202130645514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.0015159078175202012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940988120622933 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.001940988120622933, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940988120622933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08255191147327423, + 0.026537438854575157, + 0.03842649981379509, + 0.03761020302772522, + 0.04317956790328026, + 0.1829245388507843, + 0.06138915941119194, + 0.201803058385849, + 0.1705946922302246, + 0.15498289465904236 + ], + "credit_map": [ + 0.009573403745889664, + 0.00903292465955019, + 0.007264180574566126, + 0.016962526366114616, + 0.009378165006637573, + 0.011255496181547642, + 0.011913422495126724, + 0.007171789649873972, + 0.0030338133219629526, + 0.003445368492975831, + 0.0038128437008708715, + 0.0030499338172376156, + 0.002331063849851489, + 0.0033257787581533194, + 0.0032507218420505524, + 0.004287912510335445, + 0.004137691110372543, + 0.006119840778410435, + 0.004885491449385881, + 0.004055918660014868, + 0.004333860706537962, + 0.006197911221534014, + 0.005036742892116308, + 0.003659043461084366, + 0.0039041333366185427, + 0.0059429495595395565, + 0.0025472831912338734, + 0.0036019959952682257, + 0.00675568263977766, + 0.007990792393684387, + 0.004672184120863676, + 0.002195178996771574, + 0.0040434799157083035, + 0.0038081472739577293, + 0.002749198582023382, + 0.003953773528337479, + 0.005773314740508795, + 0.013656488619744778, + 0.00615461403504014, + 0.0030405488796532154, + 0.018467934802174568, + 0.030735211446881294, + 0.012089740484952927, + 0.01757957600057125, + 0.03515436500310898, + 0.040504977107048035, + 0.01637083850800991, + 0.012021888978779316, + 0.006841023452579975, + 0.01030519139021635, + 0.005142000503838062, + 0.0064326985739171505, + 0.010454051196575165, + 0.011320789344608784, + 0.006074636708945036, + 0.004818761721253395, + 0.0271347314119339, + 0.03796267881989479, + 0.012342744506895542, + 0.02097678929567337, + 0.048313774168491364, + 0.024047166109085083, + 0.016902731731534004, + 0.014122426509857178, + 0.023587556555867195, + 0.03354329615831375, + 0.011170937679708004, + 0.016213590279221535, + 0.03738198056817055, + 0.020962195470929146, + 0.013245292007923126, + 0.014489836990833282, + 0.02255496010184288, + 0.03133946657180786, + 0.008554959669709206, + 0.014425438828766346, + 0.036190688610076904, + 0.018923969939351082, + 0.013061744160950184, + 0.009931668639183044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.001940988120622933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735360249876976 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735360249876976, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735360249876976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08443673700094223, + 0.023571837693452835, + 0.03855348750948906, + 0.018943388015031815, + 0.012912343256175518, + 0.264146089553833, + 0.02393648400902748, + 0.2859593629837036, + 0.1955132782459259, + 0.052026983350515366 + ], + "credit_map": [ + 0.01142149604856968, + 0.018110228702425957, + 0.008299427106976509, + 0.00992068462073803, + 0.012241178192198277, + 0.008243639953434467, + 0.008907206356525421, + 0.007292869500815868, + 0.002601504558697343, + 0.002281766850501299, + 0.0028655475471168756, + 0.002061100909486413, + 0.002654467709362507, + 0.005984377581626177, + 0.0025299538392573595, + 0.002593118464574218, + 0.0046405065804719925, + 0.0016747505869716406, + 0.00216114753857255, + 0.0027029954362660646, + 0.002618007129058242, + 0.015093101188540459, + 0.007358279544860125, + 0.0023046990390866995, + 0.0021811325568705797, + 0.0030879380647093058, + 0.0017069501336663961, + 0.0020357368048280478, + 0.002916851080954075, + 0.003949183039367199, + 0.001915489905513823, + 0.0011501030530780554, + 0.0015155543806031346, + 0.0021321524400264025, + 0.0013913706643506885, + 0.0008665298228152096, + 0.0022437237203121185, + 0.0022519072517752647, + 0.0016051355050876737, + 0.000905968074221164, + 0.026445569470524788, + 0.051273591816425323, + 0.018594002351164818, + 0.022554665803909302, + 0.058669641613960266, + 0.04824070259928703, + 0.02139631286263466, + 0.016971608623862267, + 0.00307227554731071, + 0.0031856324058026075, + 0.0023245024494826794, + 0.0031949805561453104, + 0.0035151627380400896, + 0.004214342217892408, + 0.0022979287896305323, + 0.0021316574420779943, + 0.04086204618215561, + 0.05358287692070007, + 0.01734926737844944, + 0.028809519484639168, + 0.06953111290931702, + 0.03307104855775833, + 0.02444022335112095, + 0.018313243985176086, + 0.027001677080988884, + 0.040203288197517395, + 0.01174326241016388, + 0.018322672694921494, + 0.04565422236919403, + 0.022341184318065643, + 0.015264064073562622, + 0.014982893131673336, + 0.0073844376020133495, + 0.010461518540978432, + 0.00279656401835382, + 0.004883128684014082, + 0.01282584574073553, + 0.006339774467051029, + 0.004261080641299486, + 0.003074625274166465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.006735360249876976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures the scene, rendered in the friendly, colorful style of Pixar animation, focusing on the perfect centerpiece of a lush, mossy forest floor. The scene is bathed in warm, dappled sunlight filtering through the unseen treetops, creating softly edged shadows. A small mound amidst the emerald green moss begins to pulsate, and then a captivating fungus powerfully thrusts into view. Its creamy white stem, dotted with cheerful reddish-brown spots, rapidly elongates upward. The cap, initially a compact, vibrant pumpkin-orange sphere, gracefully unfolds in time-lapse photography, its smooth, saturated surface expanding outward to reveal its full, classic mushroom form. As it reaches its final, proud stance, the entire mushroom exhibits a subtle, endearing sway, blending into the magical forest scene with a palpable sense of life and whimsy.", + "seed": 113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_6012_seed_113_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": -0.0036464288365095853, + "vlm_logprob": -0.0036464288365095853, + "vlm_yes_no_margin": 6.2, + "vlm_token_credit": -0.0036464288365095853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005885888356715441 + ], + "opposite_logprob_tokens": [ + -7.380886077880859 + ], + "logprob_reward": -0.005885888356715441, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.005885888356715441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12153118848800659, + 0.052612561732530594, + 0.03057565726339817, + 0.0737200677394867, + 0.10278551280498505, + 0.2912558317184448, + 0.1250908523797989, + 0.11202704906463623, + 0.052297260612249374, + 0.03810393437743187 + ], + "credit_map": [ + 0.01086432859301567, + 0.02075737901031971, + 0.01017799787223339, + 0.011188163422048092, + 0.018586406484246254, + 0.016750531271100044, + 0.02252339944243431, + 0.01068299263715744, + 0.00912169087678194, + 0.010679359547793865, + 0.005057720933109522, + 0.0040328409522771835, + 0.00662697060033679, + 0.007102509960532188, + 0.006017211824655533, + 0.003974262624979019, + 0.00624106265604496, + 0.003710215911269188, + 0.00286508328281343, + 0.002258546883240342, + 0.004519124981015921, + 0.004655268043279648, + 0.003662042785435915, + 0.0026643150486052036, + 0.012765645049512386, + 0.008667058311402798, + 0.005910105537623167, + 0.007052314933389425, + 0.009753052145242691, + 0.013370247557759285, + 0.009593096561729908, + 0.0066085560247302055, + 0.014264310710132122, + 0.006906762719154358, + 0.0053293658420443535, + 0.006274477578699589, + 0.006876912899315357, + 0.029384156689047813, + 0.012064233422279358, + 0.02168530970811844, + 0.044390611350536346, + 0.043403010815382004, + 0.020031509920954704, + 0.031091749668121338, + 0.033164333552122116, + 0.03258670121431351, + 0.05909388139843941, + 0.027494102716445923, + 0.014585715718567371, + 0.006274808663874865, + 0.006574221421033144, + 0.009737693704664707, + 0.008532841689884663, + 0.04222501069307327, + 0.018804121762514114, + 0.018356461077928543, + 0.009227496571838856, + 0.007084724027663469, + 0.00354900024831295, + 0.0051012588664889336, + 0.008414181880652905, + 0.03888283297419548, + 0.008936973288655281, + 0.030830590054392815, + 0.006050620228052139, + 0.00451617268845439, + 0.004428712651133537, + 0.008364913985133171, + 0.006251961924135685, + 0.004922967869788408, + 0.011666114442050457, + 0.006095804739743471, + 0.005583324935287237, + 0.004022589419037104, + 0.004506751894950867, + 0.004019337706267834, + 0.004425544757395983, + 0.004587717819958925, + 0.0053513371385633945, + 0.0056073348969221115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.005885888356715441 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017171651124954224 + ], + "opposite_logprob_tokens": [ + -6.3767170906066895 + ], + "logprob_reward": -0.0017171651124954224, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017171651124954224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10081909596920013, + 0.06488240510225296, + 0.03563765436410904, + 0.03331386670470238, + 0.05145667865872383, + 0.15778328478336334, + 0.18030042946338654, + 0.24397769570350647, + 0.047099072486162186, + 0.08472984284162521 + ], + "credit_map": [ + 0.015247969888150692, + 0.02162053808569908, + 0.01173415407538414, + 0.004653261974453926, + 0.01054620835930109, + 0.019741836935281754, + 0.006567609496414661, + 0.010707521811127663, + 0.010056755505502224, + 0.009220910258591175, + 0.0061089531518518925, + 0.006025669164955616, + 0.00789695791900158, + 0.008852408267557621, + 0.010138841345906258, + 0.006581906694918871, + 0.003773686708882451, + 0.003871261142194271, + 0.0051051839254796505, + 0.0025560655631124973, + 0.00721265934407711, + 0.0037628239952027798, + 0.004736054688692093, + 0.004619919694960117, + 0.006683226674795151, + 0.002873535268008709, + 0.004538515117019415, + 0.00265044835396111, + 0.0047019170597195625, + 0.005469702184200287, + 0.003529940266162157, + 0.0028665787540376186, + 0.005162642803043127, + 0.0041288272477686405, + 0.00413794768974185, + 0.003956914879381657, + 0.005175363272428513, + 0.011327633634209633, + 0.007533696945756674, + 0.010033656843006611, + 0.023077230900526047, + 0.023916233330965042, + 0.010350842028856277, + 0.015687409788370132, + 0.017439676448702812, + 0.02072993665933609, + 0.030154971405863762, + 0.01642698049545288, + 0.019389266148209572, + 0.008570516481995583, + 0.006783589720726013, + 0.010958136059343815, + 0.010526568628847599, + 0.06837419420480728, + 0.021336328238248825, + 0.0343618243932724, + 0.022083275020122528, + 0.017201842740178108, + 0.010833533480763435, + 0.020961057394742966, + 0.020699549466371536, + 0.06755517423152924, + 0.03308422490954399, + 0.051559023559093475, + 0.005265037529170513, + 0.0063753193244338036, + 0.003810306079685688, + 0.004647387191653252, + 0.005521600600332022, + 0.004426175262778997, + 0.010144776664674282, + 0.006908472627401352, + 0.007033555768430233, + 0.015004743821918964, + 0.009434857405722141, + 0.009368469938635826, + 0.0069669075310230255, + 0.00698269996792078, + 0.01147676631808281, + 0.018461842089891434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.0017171651124954224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024895647075027227 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024895647075027227, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024895647075027227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1034318134188652, + 0.04479316994547844, + 0.06670964509248734, + 0.06625845283269882, + 0.02501109428703785, + 0.18974798917770386, + 0.07036610692739487, + 0.3256028890609741, + 0.07706761360168457, + 0.031011270359158516 + ], + "credit_map": [ + 0.01602218858897686, + 0.016698867082595825, + 0.005734503734856844, + 0.008499374613165855, + 0.01714380644261837, + 0.014218417927622795, + 0.01922815665602684, + 0.005886495113372803, + 0.008200894109904766, + 0.00649087643250823, + 0.0031690020114183426, + 0.0036677359603345394, + 0.005655424669384956, + 0.0066504934802651405, + 0.006660871207714081, + 0.0042978692799806595, + 0.007391279097646475, + 0.00791460182517767, + 0.003687321674078703, + 0.005255645606666803, + 0.01681947521865368, + 0.009625967592000961, + 0.010174062103033066, + 0.005841286852955818, + 0.008950209245085716, + 0.007644018158316612, + 0.005178709980100393, + 0.0065469094552099705, + 0.011428162455558777, + 0.008138584904372692, + 0.010474028065800667, + 0.00789782777428627, + 0.003378849243745208, + 0.0020144032314419746, + 0.0019301786087453365, + 0.0025593964383006096, + 0.001987556228414178, + 0.004050713963806629, + 0.00557002704590559, + 0.003519968595355749, + 0.026047302410006523, + 0.027346931397914886, + 0.011983225122094154, + 0.016060642898082733, + 0.019489699974656105, + 0.03138197585940361, + 0.03289724513888359, + 0.024540970101952553, + 0.008255967870354652, + 0.004855316132307053, + 0.006104588508605957, + 0.008704619482159615, + 0.0061748214066028595, + 0.013277583755552769, + 0.017021488398313522, + 0.005971716716885567, + 0.038155101239681244, + 0.03281378000974655, + 0.022110428661108017, + 0.04413089528679848, + 0.038947753608226776, + 0.039194103330373764, + 0.07739180326461792, + 0.03285900875926018, + 0.008963936939835548, + 0.010050195269286633, + 0.005569536704570055, + 0.008038616739213467, + 0.009715752676129341, + 0.0068022264167666435, + 0.015517100691795349, + 0.012410248629748821, + 0.003878901246935129, + 0.0028741711284965277, + 0.0023879047948867083, + 0.003924054559320211, + 0.004157752729952335, + 0.005008685868233442, + 0.005802913103252649, + 0.0029768841341137886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.0024895647075027227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195806708186865 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195806708186865, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195806708186865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23184920847415924, + 0.09046255052089691, + 0.10050380229949951, + 0.029950816184282303, + 0.033901065587997437, + 0.1664343625307083, + 0.0374101847410202, + 0.17133620381355286, + 0.05096517875790596, + 0.08718668669462204 + ], + "credit_map": [ + 0.021825948730111122, + 0.028430424630641937, + 0.011126598343253136, + 0.022703176364302635, + 0.038032688200473785, + 0.03966188430786133, + 0.056729789823293686, + 0.013338690623641014, + 0.009916337206959724, + 0.015095542185008526, + 0.005043928511440754, + 0.007392248138785362, + 0.01527806743979454, + 0.011424937285482883, + 0.017669903114438057, + 0.008641577325761318, + 0.013280784711241722, + 0.011248881928622723, + 0.004887478891760111, + 0.006051742006093264, + 0.027206605300307274, + 0.013627436943352222, + 0.0158785879611969, + 0.0083222771063447, + 0.005040106363594532, + 0.0036341806408017874, + 0.0016750111244618893, + 0.003076376859098673, + 0.00410059466958046, + 0.004390195477753878, + 0.005096814129501581, + 0.0029375345911830664, + 0.004400751553475857, + 0.003930999897420406, + 0.0031507781241089106, + 0.003749104682356119, + 0.0030204118229448795, + 0.005040305200964212, + 0.0068374923430383205, + 0.0037712182383984327, + 0.021078957244753838, + 0.016165105625987053, + 0.006761256605386734, + 0.00997013971209526, + 0.012055383063852787, + 0.04787931218743324, + 0.019046256318688393, + 0.033477939665317535, + 0.004291524179279804, + 0.00391309754922986, + 0.0017895643832162023, + 0.003254578448832035, + 0.0028461951296776533, + 0.010055753402411938, + 0.005609400570392609, + 0.005650071427226067, + 0.014704924076795578, + 0.009207421913743019, + 0.00642371503636241, + 0.010387877002358437, + 0.01192590594291687, + 0.056283872574567795, + 0.02151133492588997, + 0.0408911257982254, + 0.005798725411295891, + 0.005912146996706724, + 0.0034409635700285435, + 0.006366563029587269, + 0.006928927730768919, + 0.0064830719493329525, + 0.008960618637502193, + 0.007074157707393169, + 0.006520318333059549, + 0.018661441281437874, + 0.004178149625658989, + 0.005621891934424639, + 0.012903757393360138, + 0.01058577187359333, + 0.01339026726782322, + 0.01532508060336113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.002195806708186865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943719297647476 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943719297647476, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943719297647476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1465195268392563, + 0.06261558830738068, + 0.0514170303940773, + 0.032657403498888016, + 0.026978138834238052, + 0.3316271901130676, + 0.12938985228538513, + 0.09813252836465836, + 0.053967393934726715, + 0.06669537723064423 + ], + "credit_map": [ + 0.025297701358795166, + 0.023048264905810356, + 0.007021394092589617, + 0.010902001522481441, + 0.023025184869766235, + 0.023233072832226753, + 0.025563228875398636, + 0.008428667671978474, + 0.011043081060051918, + 0.010174191556870937, + 0.004957813769578934, + 0.0046141743659973145, + 0.0071657090447843075, + 0.008795278146862984, + 0.009354094974696636, + 0.006511244922876358, + 0.007068349979817867, + 0.005507002584636211, + 0.0038875590544193983, + 0.0037177703343331814, + 0.011682226322591305, + 0.007040014490485191, + 0.008055971935391426, + 0.0044581349939107895, + 0.0055762105621397495, + 0.004772869870066643, + 0.004522308707237244, + 0.0027709645219147205, + 0.0030439214315265417, + 0.004588197451084852, + 0.004138269927352667, + 0.003244661260396242, + 0.0036218250170350075, + 0.002640867605805397, + 0.002856489270925522, + 0.0027872638311237097, + 0.0021780335810035467, + 0.004308639094233513, + 0.005268832203000784, + 0.0033161877654492855, + 0.044095080345869064, + 0.04437753185629845, + 0.017690449953079224, + 0.02684078738093376, + 0.03348517045378685, + 0.06498534232378006, + 0.052039772272109985, + 0.04811304807662964, + 0.01387668214738369, + 0.0057833874598145485, + 0.0036153020337224007, + 0.004718484356999397, + 0.00859018974006176, + 0.052118513733148575, + 0.011715203523635864, + 0.028972074389457703, + 0.009231733158230782, + 0.0038210111670196056, + 0.003358862129971385, + 0.00532887689769268, + 0.004611821845173836, + 0.03464877977967262, + 0.012141101993620396, + 0.024990342557430267, + 0.006462987046688795, + 0.006418927572667599, + 0.004178149159997702, + 0.006575975567102432, + 0.006173423957079649, + 0.005329208914190531, + 0.011877152137458324, + 0.0069515700452029705, + 0.008270380087196827, + 0.007006334140896797, + 0.008493548259139061, + 0.009840385988354683, + 0.007061497308313847, + 0.007027390878647566, + 0.010452432557940483, + 0.008543412201106548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.005943719297647476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle shot captures the scene, rendered in the friendly, colorful style of Pixar animation, focusing on the perfect centerpiece of a lush, mossy forest floor. The scene is bathed in warm, dappled sunlight filtering through the unseen treetops, creating softly edged shadows. A small mound amidst the emerald green moss begins to pulsate, and then a captivating fungus powerfully thrusts into view. Its creamy white stem, dotted with cheerful reddish-brown spots, rapidly elongates upward. The cap, initially a compact, vibrant pumpkin-orange sphere, gracefully unfolds in time-lapse photography, its smooth, saturated surface expanding outward to reveal its full, classic mushroom form. As it reaches its final, proud stance, the entire mushroom exhibits a subtle, endearing sway, blending into the magical forest scene with a palpable sense of life and whimsy.", + "seed": 114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_13/index_6012_seed_114_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": -0.005036546289920807, + "vlm_logprob": -0.005036546289920807, + "vlm_yes_no_margin": 5.9, + "vlm_token_credit": -0.005036546289920807, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004247572273015976 + ], + "opposite_logprob_tokens": [ + -7.379247665405273 + ], + "logprob_reward": -0.004247572273015976, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.004247572273015976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27395498752593994, + 0.08666492998600006, + 0.043089885264635086, + 0.1455184370279312, + 0.057779330760240555, + 0.03940398618578911, + 0.10433610528707504, + 0.05058083310723305, + 0.12752193212509155, + 0.0711495578289032 + ], + "credit_map": [ + 0.02208513393998146, + 0.022638486698269844, + 0.0188149306923151, + 0.016140496358275414, + 0.07903790473937988, + 0.07075794786214828, + 0.018946027383208275, + 0.02553408034145832, + 0.00801636278629303, + 0.005525487940758467, + 0.009135279804468155, + 0.011909076943993568, + 0.006385005544871092, + 0.015137625858187675, + 0.02309758961200714, + 0.007458503358066082, + 0.004305132199078798, + 0.0035003540106117725, + 0.005917235743254423, + 0.005348603241145611, + 0.002945533487945795, + 0.007619432173669338, + 0.00932458508759737, + 0.004129009786993265, + 0.01432040985673666, + 0.017828362062573433, + 0.014809481799602509, + 0.01630444824695587, + 0.015870485454797745, + 0.01962895691394806, + 0.030757255852222443, + 0.01599903218448162, + 0.0065739150159060955, + 0.006806548684835434, + 0.008248797617852688, + 0.004905751906335354, + 0.005525035317987204, + 0.008412224240601063, + 0.010753443464636803, + 0.006553612649440765, + 0.005242062732577324, + 0.00524424621835351, + 0.005293192807585001, + 0.0035363745409995317, + 0.00562279811128974, + 0.003755588550120592, + 0.00515024596825242, + 0.0055594779551029205, + 0.02264745905995369, + 0.00847496185451746, + 0.010273830033838749, + 0.009862789884209633, + 0.008945353329181671, + 0.011986003257334232, + 0.023375442251563072, + 0.00877026654779911, + 0.009138688445091248, + 0.006119015160948038, + 0.006546962074935436, + 0.00468529062345624, + 0.005468099843710661, + 0.004984651226550341, + 0.008195873349905014, + 0.00544225238263607, + 0.02364167757332325, + 0.02673853188753128, + 0.014009261503815651, + 0.00956179853528738, + 0.014499679207801819, + 0.01323117595165968, + 0.013223863206803799, + 0.012615947984158993, + 0.008360867388546467, + 0.008120022714138031, + 0.016392676159739494, + 0.0059047178365290165, + 0.006246436852961779, + 0.006512513849884272, + 0.010624277405440807, + 0.008988043293356895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.004247572273015976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017199022695422173 + ], + "opposite_logprob_tokens": [ + -6.376719951629639 + ], + "logprob_reward": -0.0017199022695422173, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017199022695422173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.274490088224411, + 0.12440339475870132, + 0.06972827762365341, + 0.05366988480091095, + 0.08392030745744705, + 0.05540917068719864, + 0.0806552916765213, + 0.05005976930260658, + 0.10360175371170044, + 0.10406208783388138 + ], + "credit_map": [ + 0.034907735884189606, + 0.01752074994146824, + 0.029517948627471924, + 0.015617677941918373, + 0.058481745421886444, + 0.06824857741594315, + 0.024422140792012215, + 0.02577349916100502, + 0.012354020960628986, + 0.008834385313093662, + 0.010825070552527905, + 0.016842437908053398, + 0.00886059831827879, + 0.024020327255129814, + 0.030672239139676094, + 0.01199430599808693, + 0.009123913012444973, + 0.0055474708788096905, + 0.011673638597130775, + 0.007959871552884579, + 0.005443336442112923, + 0.01013620849698782, + 0.012361275963485241, + 0.007482554763555527, + 0.006894061807543039, + 0.006162950303405523, + 0.006748508661985397, + 0.0059592449106276035, + 0.005294775124639273, + 0.0062080733478069305, + 0.009354153648018837, + 0.007048117928206921, + 0.007949025370180607, + 0.02301863022148609, + 0.010584144853055477, + 0.004762584809213877, + 0.011976322159171104, + 0.008256131783127785, + 0.009998471476137638, + 0.007374998647719622, + 0.005462260916829109, + 0.006785350386053324, + 0.007814853452146053, + 0.005584662314504385, + 0.005645155441015959, + 0.006854057312011719, + 0.011835971847176552, + 0.005426862742751837, + 0.006720204837620258, + 0.011223489418625832, + 0.010077579878270626, + 0.008903670124709606, + 0.009705668315291405, + 0.008941569365561008, + 0.01867459900677204, + 0.006408511660993099, + 0.0049507394433021545, + 0.0069525958970189095, + 0.005628831218928099, + 0.005954093299806118, + 0.004700665362179279, + 0.004868496675044298, + 0.012873170897364616, + 0.004131178837269545, + 0.017572037875652313, + 0.01230629626661539, + 0.01663755066692829, + 0.009964617900550365, + 0.008378870785236359, + 0.012057159096002579, + 0.01607149839401245, + 0.010613721795380116, + 0.009950580075383186, + 0.03218119591474533, + 0.008114054799079895, + 0.0063490839675068855, + 0.016498761251568794, + 0.011050212197005749, + 0.00985738541930914, + 0.010060817934572697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0017199022695422173 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199375070631504 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.002199375070631504, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199375070631504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16167044639587402, + 0.09371811896562576, + 0.08904220908880234, + 0.11569752544164658, + 0.049619872123003006, + 0.08088669180870056, + 0.07323474436998367, + 0.09086224436759949, + 0.1278519630432129, + 0.11741619557142258 + ], + "credit_map": [ + 0.02619853988289833, + 0.015294275246560574, + 0.016054876148700714, + 0.016649987548589706, + 0.018054600805044174, + 0.019332613795995712, + 0.03508070483803749, + 0.015004847198724747, + 0.015814131125807762, + 0.005501990206539631, + 0.009259125217795372, + 0.013345375657081604, + 0.007430911995470524, + 0.014415754936635494, + 0.017736133188009262, + 0.01021469384431839, + 0.011325552128255367, + 0.006491273641586304, + 0.011684696190059185, + 0.01136055402457714, + 0.005385328084230423, + 0.01244701724499464, + 0.020776085555553436, + 0.009571700356900692, + 0.013957186602056026, + 0.013487598858773708, + 0.014768233522772789, + 0.011636543087661266, + 0.012098582461476326, + 0.019750850275158882, + 0.017460821196436882, + 0.012537705712020397, + 0.004843853879719973, + 0.010308366268873215, + 0.006170832086354494, + 0.0038092017639428377, + 0.0061704362742602825, + 0.00478340545669198, + 0.008075947873294353, + 0.005457830615341663, + 0.009603871032595634, + 0.012160037644207478, + 0.007028684951364994, + 0.008415114134550095, + 0.009298133663833141, + 0.007449882570654154, + 0.02097666636109352, + 0.005954297725111246, + 0.009234901517629623, + 0.008645566180348396, + 0.00832928903400898, + 0.0068716853857040405, + 0.008329536765813828, + 0.006912233307957649, + 0.018375137820839882, + 0.006536398082971573, + 0.00811806321144104, + 0.010501246899366379, + 0.008266725577414036, + 0.012345464900135994, + 0.007753011770546436, + 0.00812143087387085, + 0.02746468223631382, + 0.008291623555123806, + 0.009188111871480942, + 0.031050939112901688, + 0.021630069240927696, + 0.011457549408078194, + 0.01602224074304104, + 0.007855934090912342, + 0.022712955251336098, + 0.00793415680527687, + 0.00954633392393589, + 0.017634419724345207, + 0.022037342190742493, + 0.012598887085914612, + 0.011814401485025883, + 0.009555759839713573, + 0.024150386452674866, + 0.0100786704570055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.002199375070631504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946681834757328 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946681834757328, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946681834757328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18842501938343048, + 0.1371871829032898, + 0.05464620143175125, + 0.07988660782575607, + 0.1345774084329605, + 0.06659726053476334, + 0.04382247105240822, + 0.04905194789171219, + 0.11776789277791977, + 0.12803800404071808 + ], + "credit_map": [ + 0.015148867852985859, + 0.015505066141486168, + 0.012180319055914879, + 0.018971828743815422, + 0.04746299609541893, + 0.03595535084605217, + 0.027542656287550926, + 0.015657935291528702, + 0.01354795042425394, + 0.010480080731213093, + 0.014820531941950321, + 0.017051585018634796, + 0.01072229165583849, + 0.022504180669784546, + 0.035332631319761276, + 0.012727931141853333, + 0.007440176326781511, + 0.004066657740622759, + 0.009204856120049953, + 0.006188584491610527, + 0.002946204040199518, + 0.009969508275389671, + 0.007963049225509167, + 0.006867159623652697, + 0.009033280424773693, + 0.011334679089486599, + 0.009459666907787323, + 0.007362245582044125, + 0.007186560425907373, + 0.01673237979412079, + 0.009985563345253468, + 0.008792228065431118, + 0.01278997864574194, + 0.02705899253487587, + 0.021729422733187675, + 0.010305553674697876, + 0.016164788976311684, + 0.012189682573080063, + 0.02485966868698597, + 0.009479322470724583, + 0.008701888844370842, + 0.01084402296692133, + 0.006264601834118366, + 0.006036775186657906, + 0.007928536273539066, + 0.006103502120822668, + 0.015438171103596687, + 0.005279761739075184, + 0.004327114671468735, + 0.009324259124696255, + 0.0038575923535972834, + 0.004132646601647139, + 0.007376824039965868, + 0.0035078879445791245, + 0.007890243083238602, + 0.0034059036988765, + 0.005754363723099232, + 0.007676287088543177, + 0.0032558797392994165, + 0.005564381368458271, + 0.0059213535860180855, + 0.004591687116771936, + 0.012432210147380829, + 0.0038557846564799547, + 0.017906896770000458, + 0.02594592049717903, + 0.015566029585897923, + 0.009354236535727978, + 0.013023244217038155, + 0.01123722456395626, + 0.015056733973324299, + 0.009677606634795666, + 0.009628319181501865, + 0.04138144850730896, + 0.030740024521946907, + 0.006173504516482353, + 0.017550205811858177, + 0.005781666375696659, + 0.008041576482355595, + 0.00874125212430954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005946681834757328 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a friendly, colorful Pixar animation style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a lush, mossy forest floor with a small mound of emerald green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a fungus emerge from the mound with a creamy white stem dotted with reddish-brown spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mushroom cap start as a compact pumpkin-orange sphere and unfold into a full mushroom form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069200001657009 + ], + "opposite_logprob_tokens": [ + -4.511069297790527 + ], + "logprob_reward": -0.011069200001657009, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069200001657009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17848928272724152, + 0.08769457787275314, + 0.03974383324384689, + 0.1509043574333191, + 0.23637425899505615, + 0.08175317198038101, + 0.053417399525642395, + 0.03124239854514599, + 0.06625718623399734, + 0.07412353157997131 + ], + "credit_map": [ + 0.017452877014875412, + 0.015034317970275879, + 0.016401056200265884, + 0.010764388367533684, + 0.04518526792526245, + 0.037940435111522675, + 0.019340693950653076, + 0.0163702592253685, + 0.013070524670183659, + 0.006494493689388037, + 0.00900760293006897, + 0.00974840484559536, + 0.0065366788767278194, + 0.011965440586209297, + 0.02349666878581047, + 0.007374767679721117, + 0.003966950811445713, + 0.0038348978850990534, + 0.005026599392294884, + 0.004713239148259163, + 0.004169234540313482, + 0.006079455837607384, + 0.006782017648220062, + 0.005171441473066807, + 0.01558157429099083, + 0.020126188173890114, + 0.017712203785777092, + 0.01332008559256792, + 0.013576706871390343, + 0.03406320512294769, + 0.022079411894083023, + 0.014444984495639801, + 0.01647692546248436, + 0.04218631610274315, + 0.03985424339771271, + 0.019537586718797684, + 0.023297147825360298, + 0.02545197308063507, + 0.04538466036319733, + 0.0241854190826416, + 0.008015003986656666, + 0.013813039287924767, + 0.008584784343838692, + 0.007555417250841856, + 0.010830463841557503, + 0.008709054440259933, + 0.016726333647966385, + 0.00751908216625452, + 0.01400664635002613, + 0.007572340779006481, + 0.0067993756383657455, + 0.003856170456856489, + 0.004546422511339188, + 0.00452809501439333, + 0.00710144080221653, + 0.005006909370422363, + 0.004885813221335411, + 0.006767562590539455, + 0.004131770692765713, + 0.0025528522673994303, + 0.004187718965113163, + 0.002500122645869851, + 0.0033487696200609207, + 0.002867789240553975, + 0.007603364996612072, + 0.02197835221886635, + 0.00873398408293724, + 0.003407828975468874, + 0.010912465862929821, + 0.005116441287100315, + 0.00367707759141922, + 0.004827677272260189, + 0.005370976869016886, + 0.027837134897708893, + 0.013383390381932259, + 0.0033678284380584955, + 0.012163832783699036, + 0.003977891057729721, + 0.004211034160107374, + 0.0038114525377750397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.011069200001657009 + } + ], + "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_14/index_tensor([4429, 4429])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_14/index_tensor([4429, 4429])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d8c945b44c70b1ec9d2ff1f06db28801f9e72029 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_14/index_tensor([4429, 4429])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, orbiting medium shot captures a female figure skater in poetic slow motion, executing a triple axel at the pinnacle of her performance. The camera moves fluidly around her, synchronized with her rotation, locking the viewer's gaze firmly on her. She wears a deep blue competition suit, studded with countless shimmering crystals. The Lycra material clings to her toned and graceful figure. As she spins at high speed, her dark brown hair, tightly gathered in a bun, escapes, tracing graceful arcs through the air. In slow motion, she is at the apex of her spin, her arms clasped tightly against her chest to maximize angular velocity. Her posture is impeccable, demonstrating supreme focus and control. The intense spotlight from the stadium's ceiling casts a bright, dramatic rim light across her figure, completely separating her from the background. This light not only emphasizes the contours of her muscles but also makes each crystal on her suit radiate a star-like brilliance. Behind her, the vast auditorium blurred into a soft, shimmering backdrop of varying shades of color, like a flowing Impressionist painting. The entire scene, reflected off the mirror-like surface of the ice, was imbued with a tension of power and beauty, creating a breathtaking, almost sacred atmosphere of competition.", + "seed": 16692, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_14/index_4429_seed_16692_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": -0.1347373732365668, + "vlm_logprob": -0.1347373732365668, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.1347373732365668, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014761041849851608 + ], + "opposite_logprob_tokens": [ + -4.389760971069336 + ], + "logprob_reward": -0.014761041849851608, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014761041849851608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14330916106700897, + 0.15332764387130737, + 0.08870911598205566, + 0.03511711210012436, + 0.04762878641486168, + 0.2800215780735016, + 0.09949268400669098, + 0.0902363657951355, + 0.03253389149904251, + 0.02962365187704563 + ], + "credit_map": [ + 0.014079411514103413, + 0.014846878126263618, + 0.025228377431631088, + 0.01809680089354515, + 0.022454120218753815, + 0.013088054023683071, + 0.022737713530659676, + 0.012777799740433693, + 0.029221050441265106, + 0.01615748181939125, + 0.035643115639686584, + 0.0098723154515028, + 0.008705982007086277, + 0.01500418595969677, + 0.011287069879472256, + 0.02743644081056118, + 0.0086591811850667, + 0.011360922828316689, + 0.010528120212256908, + 0.012794381007552147, + 0.012640420347452164, + 0.01179379690438509, + 0.013308392837643623, + 0.007623900193721056, + 0.0034872957039624453, + 0.004079779144376516, + 0.0040688845328986645, + 0.005548684857785702, + 0.0032708460930734873, + 0.004061905667185783, + 0.005593298934400082, + 0.005006419029086828, + 0.005313906352967024, + 0.0034326971508562565, + 0.005253824405372143, + 0.004158994182944298, + 0.00786906573921442, + 0.00650431215763092, + 0.010037951171398163, + 0.00505803432315588, + 0.06778464466333389, + 0.027037853375077248, + 0.10058056563138962, + 0.023175816982984543, + 0.009685136377811432, + 0.01911466382443905, + 0.01309366524219513, + 0.019549231976270676, + 0.009208193980157375, + 0.01275296788662672, + 0.009659969247877598, + 0.011127147823572159, + 0.01303956750780344, + 0.014028237201273441, + 0.012938251718878746, + 0.016738347709178925, + 0.018289973959326744, + 0.010848701000213623, + 0.009701298549771309, + 0.005848262459039688, + 0.0059990608133375645, + 0.02742875926196575, + 0.00581494253128767, + 0.006305377464741468, + 0.003476562676951289, + 0.004639468155801296, + 0.004169919528067112, + 0.004173758439719677, + 0.004392270930111408, + 0.0036146806087344885, + 0.004547863267362118, + 0.003519368125125766, + 0.00367834628559649, + 0.0026321972254663706, + 0.005622627679258585, + 0.0036926367320120335, + 0.0026977607049047947, + 0.005070410203188658, + 0.0028552901931107044, + 0.003374381922185421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.014761041849851608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761969685554504 + ], + "opposite_logprob_tokens": [ + -0.8261969685554504 + ], + "logprob_reward": -0.5761969685554504, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761969685554504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11230428516864777, + 0.12201102823019028, + 0.2588209807872772, + 0.07405762374401093, + 0.056555893272161484, + 0.05212033912539482, + 0.06626339256763458, + 0.05102197080850601, + 0.15777580440044403, + 0.0490686409175396 + ], + "credit_map": [ + 0.009133782237768173, + 0.011245944537222385, + 0.017751548439264297, + 0.015257474035024643, + 0.0148954838514328, + 0.015475590713322163, + 0.018155893310904503, + 0.010388561524450779, + 0.018568193539977074, + 0.008801350370049477, + 0.026908937841653824, + 0.008794402703642845, + 0.010910345241427422, + 0.01014066394418478, + 0.017927244305610657, + 0.019959889352321625, + 0.019353743642568588, + 0.031535226851701736, + 0.038616444915533066, + 0.058411505073308945, + 0.02870285138487816, + 0.03636276721954346, + 0.022474095225334167, + 0.02336435578763485, + 0.01051353570073843, + 0.009034205228090286, + 0.01075551938265562, + 0.010518224909901619, + 0.008167671971023083, + 0.006841427180916071, + 0.010951392352581024, + 0.007275642827153206, + 0.004961562808603048, + 0.0041662342846393585, + 0.005901950877159834, + 0.00533221336081624, + 0.012856231071054935, + 0.007312607951462269, + 0.007950576953589916, + 0.00807451456785202, + 0.008741366676986217, + 0.006019501015543938, + 0.007859304547309875, + 0.006484237965196371, + 0.004261107183992863, + 0.008268941193819046, + 0.005191362928599119, + 0.005294520407915115, + 0.00941956415772438, + 0.015451069921255112, + 0.0075727845542132854, + 0.004567421972751617, + 0.006562685128301382, + 0.009383476339280605, + 0.0068343146704137325, + 0.006472081411629915, + 0.007204676978290081, + 0.007186684757471085, + 0.0076018511317670345, + 0.00564996525645256, + 0.006256280466914177, + 0.005992867983877659, + 0.005872196983546019, + 0.005257449112832546, + 0.016849476844072342, + 0.019158819690346718, + 0.018646907061338425, + 0.019769612699747086, + 0.01839258149266243, + 0.026638001203536987, + 0.02437584288418293, + 0.013944572769105434, + 0.0054265777580440044, + 0.003927155863493681, + 0.005184297915548086, + 0.0057777645997703075, + 0.006667713634669781, + 0.0065059480257332325, + 0.009048085659742355, + 0.006531091872602701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.5761969685554504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043087150901556015 + ], + "opposite_logprob_tokens": [ + -3.1680872440338135 + ], + "logprob_reward": -0.043087150901556015, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043087150901556015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1531176120042801, + 0.07901355624198914, + 0.21810349822044373, + 0.11505729705095291, + 0.045304909348487854, + 0.18161828815937042, + 0.04634235426783562, + 0.061720170080661774, + 0.060005735605955124, + 0.03971663489937782 + ], + "credit_map": [ + 0.016221873462200165, + 0.015456249006092548, + 0.04385049268603325, + 0.014477682299911976, + 0.01552416943013668, + 0.02146587148308754, + 0.010186197236180305, + 0.015935078263282776, + 0.01283089816570282, + 0.006920292507857084, + 0.014509234577417374, + 0.009333478286862373, + 0.006218916270881891, + 0.0061337570659816265, + 0.0063776602037250996, + 0.01668930985033512, + 0.01806403324007988, + 0.02807820960879326, + 0.030663877725601196, + 0.06086277216672897, + 0.01344598364084959, + 0.03273029997944832, + 0.013545668683946133, + 0.02071264386177063, + 0.010386909358203411, + 0.012533326633274555, + 0.014705638401210308, + 0.018254783004522324, + 0.012051168829202652, + 0.011941112577915192, + 0.02735425904393196, + 0.007830086164176464, + 0.005136169958859682, + 0.005673825740814209, + 0.005593401845544577, + 0.006467527709901333, + 0.0057768491096794605, + 0.005457476247102022, + 0.005886489059776068, + 0.005313171073794365, + 0.032583631575107574, + 0.02331547439098358, + 0.05562616512179375, + 0.012084215879440308, + 0.015360677614808083, + 0.013785436749458313, + 0.010453197173774242, + 0.018409475684165955, + 0.006287256721407175, + 0.007287443615496159, + 0.004372850991785526, + 0.004241399932652712, + 0.004332302138209343, + 0.005767912603914738, + 0.006480769254267216, + 0.007572414353489876, + 0.005881109274923801, + 0.01041873823851347, + 0.008764254860579967, + 0.004557228181511164, + 0.00827758014202118, + 0.013805425725877285, + 0.005414809565991163, + 0.004601019900292158, + 0.0057693044655025005, + 0.006767421495169401, + 0.00774540426209569, + 0.007485172711312771, + 0.010966146364808083, + 0.008292626589536667, + 0.008699910715222359, + 0.004279743880033493, + 0.0039570024237036705, + 0.0028548764530569315, + 0.007134969346225262, + 0.00550031429156661, + 0.005158292595297098, + 0.005643858574330807, + 0.0035257928539067507, + 0.005941521842032671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.043087150901556015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947985220700502 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005947985220700502, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947985220700502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09932728111743927, + 0.2254679799079895, + 0.2099800407886505, + 0.07177052646875381, + 0.07566329091787338, + 0.08990901708602905, + 0.08795681595802307, + 0.08023453503847122, + 0.026896251365542412, + 0.032794319093227386 + ], + "credit_map": [ + 0.009494663216173649, + 0.0060572270303964615, + 0.026783471927046776, + 0.01031696517020464, + 0.006423139479011297, + 0.01637972891330719, + 0.012417323887348175, + 0.011454766616225243, + 0.03773602098226547, + 0.01988760381937027, + 0.046553291380405426, + 0.019370343536138535, + 0.014122392050921917, + 0.020677246153354645, + 0.013730155304074287, + 0.05339093506336212, + 0.020589817315340042, + 0.028446964919567108, + 0.02739349566400051, + 0.059401340782642365, + 0.013051589950919151, + 0.03021847829222679, + 0.015469571575522423, + 0.015408768318593502, + 0.007918118499219418, + 0.008219834417104721, + 0.010657238774001598, + 0.013188659213483334, + 0.006963466759771109, + 0.006956050638109446, + 0.009853136725723743, + 0.008014025166630745, + 0.010314115323126316, + 0.00675596110522747, + 0.014011829160153866, + 0.007657680194824934, + 0.00541351456195116, + 0.0089905159547925, + 0.007939252071082592, + 0.014580417424440384, + 0.018628505989909172, + 0.01760217361152172, + 0.00714312307536602, + 0.008696889504790306, + 0.008842960931360722, + 0.009492059238255024, + 0.007916159927845001, + 0.011587139219045639, + 0.014868882484734058, + 0.016424691304564476, + 0.010823321528732777, + 0.007540920749306679, + 0.008880083449184895, + 0.010538401082158089, + 0.00915608648210764, + 0.009724426083266735, + 0.019311530515551567, + 0.010089099407196045, + 0.010326052084565163, + 0.0036702801007777452, + 0.003783033462241292, + 0.025062667205929756, + 0.0026146385353058577, + 0.0053772348910570145, + 0.0038428502157330513, + 0.003260013647377491, + 0.002887134440243244, + 0.002895055338740349, + 0.0025665045250207186, + 0.004933902993798256, + 0.003583903657272458, + 0.002926887245848775, + 0.0050210789777338505, + 0.00254017673432827, + 0.006308017764240503, + 0.003370639868080616, + 0.002407239517197013, + 0.004984281957149506, + 0.0022343213204294443, + 0.0059285638853907585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.005947985220700502 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033693719655275345 + ], + "opposite_logprob_tokens": [ + -3.408693790435791 + ], + "logprob_reward": -0.033693719655275345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033693719655275345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10768492519855499, + 0.0606694370508194, + 0.09214615821838379, + 0.0781860277056694, + 0.036706939339637756, + 0.45459413528442383, + 0.04595467075705528, + 0.04048639535903931, + 0.056434374302625656, + 0.02713693492114544 + ], + "credit_map": [ + 0.008949324488639832, + 0.010111453011631966, + 0.02764003351330757, + 0.010816161520779133, + 0.01049363799393177, + 0.017390524968504906, + 0.014023255556821823, + 0.008260530419647694, + 0.008776682429015636, + 0.006208913400769234, + 0.011179733090102673, + 0.0046095591969788074, + 0.00510421022772789, + 0.005163108929991722, + 0.007985755801200867, + 0.011641471646726131, + 0.009064235724508762, + 0.012441743165254593, + 0.011343073099851608, + 0.024382704868912697, + 0.006492447108030319, + 0.013555441983044147, + 0.007880273275077343, + 0.006986232474446297, + 0.007650020997971296, + 0.007906408049166203, + 0.010263386182487011, + 0.017086902633309364, + 0.00968264415860176, + 0.006170887500047684, + 0.014000661671161652, + 0.005425110459327698, + 0.0034149165730923414, + 0.003848023246973753, + 0.0044139898382127285, + 0.005622366908937693, + 0.00390631565824151, + 0.004394178744405508, + 0.004432551562786102, + 0.006674594711512327, + 0.1216234415769577, + 0.028593452647328377, + 0.15242570638656616, + 0.03732439503073692, + 0.020829783752560616, + 0.026002468541264534, + 0.02331053465604782, + 0.04448433592915535, + 0.008047710172832012, + 0.008365757763385773, + 0.004545383155345917, + 0.00424191216006875, + 0.005055213812738657, + 0.0048826890997588634, + 0.005807293113321066, + 0.005008706822991371, + 0.009720916859805584, + 0.005438519641757011, + 0.003515835152938962, + 0.001413091435097158, + 0.002439947798848152, + 0.013970947824418545, + 0.0016446170629933476, + 0.002342513995245099, + 0.006009234581142664, + 0.00814689602702856, + 0.007632758934050798, + 0.007318997755646706, + 0.006578153930604458, + 0.007550109643489122, + 0.007954537868499756, + 0.005243681836873293, + 0.0038981104735285044, + 0.0020895889028906822, + 0.005644060205668211, + 0.0025517286267131567, + 0.0020564929582178593, + 0.004021678119897842, + 0.0019584433175623417, + 0.004916829522699118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.033693719655275345 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, orbiting medium shot captures a female figure skater in poetic slow motion, executing a triple axel at the pinnacle of her performance. The camera moves fluidly around her, synchronized with her rotation, locking the viewer's gaze firmly on her. She wears a deep blue competition suit, studded with countless shimmering crystals. The Lycra material clings to her toned and graceful figure. As she spins at high speed, her dark brown hair, tightly gathered in a bun, escapes, tracing graceful arcs through the air. In slow motion, she is at the apex of her spin, her arms clasped tightly against her chest to maximize angular velocity. Her posture is impeccable, demonstrating supreme focus and control. The intense spotlight from the stadium's ceiling casts a bright, dramatic rim light across her figure, completely separating her from the background. This light not only emphasizes the contours of her muscles but also makes each crystal on her suit radiate a star-like brilliance. Behind her, the vast auditorium blurred into a soft, shimmering backdrop of varying shades of color, like a flowing Impressionist painting. The entire scene, reflected off the mirror-like surface of the ice, was imbued with a tension of power and beauty, creating a breathtaking, almost sacred atmosphere of competition.", + "seed": 16693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_14/index_4429_seed_16693_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": -0.09238850027322769, + "vlm_logprob": -0.09238850027322769, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.09238850027322769, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009254653006792068 + ], + "opposite_logprob_tokens": [ + -5.009254455566406 + ], + "logprob_reward": -0.009254653006792068, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009254653006792068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15751913189888, + 0.04677238315343857, + 0.045913733541965485, + 0.1174285039305687, + 0.06292887032032013, + 0.11510080844163895, + 0.1138596162199974, + 0.14829960465431213, + 0.13078394532203674, + 0.06139334663748741 + ], + "credit_map": [ + 0.020203029736876488, + 0.014805884100496769, + 0.018076885491609573, + 0.019846368581056595, + 0.013601928018033504, + 0.01762101612985134, + 0.01623379811644554, + 0.03713022544980049, + 0.007196487858891487, + 0.005247932858765125, + 0.005498445592820644, + 0.004037963226437569, + 0.0065324753522872925, + 0.005958103109151125, + 0.005217366851866245, + 0.00708360830321908, + 0.00801609456539154, + 0.00583443371579051, + 0.006173025816679001, + 0.0046665905974805355, + 0.006259494461119175, + 0.005662568844854832, + 0.0037043204065412283, + 0.005597205832600594, + 0.011912877671420574, + 0.00755859212949872, + 0.03184735029935837, + 0.011324579827487469, + 0.007325867656618357, + 0.01830396242439747, + 0.009355826303362846, + 0.019799450412392616, + 0.00569455511868, + 0.006924121640622616, + 0.004777627065777779, + 0.006329459138214588, + 0.008642381057143211, + 0.010457988828420639, + 0.007921043783426285, + 0.012181694619357586, + 0.025946062058210373, + 0.012227291241288185, + 0.02044885791838169, + 0.007383767981082201, + 0.010423131287097931, + 0.01603424921631813, + 0.010942192748188972, + 0.011695257388055325, + 0.020988525822758675, + 0.012718947604298592, + 0.017734892666339874, + 0.006835890933871269, + 0.010079707019031048, + 0.011880683712661266, + 0.012051568366587162, + 0.021569399163126945, + 0.018573040142655373, + 0.008629302494227886, + 0.03554396703839302, + 0.009189539588987827, + 0.011054610833525658, + 0.04148796200752258, + 0.009803120978176594, + 0.014018066227436066, + 0.01680496707558632, + 0.01477959007024765, + 0.015674587339162827, + 0.009841562248766422, + 0.011859177611768246, + 0.024469759315252304, + 0.021088216453790665, + 0.016266074031591415, + 0.007023971993476152, + 0.006035596132278442, + 0.009330525062978268, + 0.005432791076600552, + 0.005797913763672113, + 0.010651555843651295, + 0.0062963226810097694, + 0.010824671015143394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.009254653006792068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871183395385742 + ], + "opposite_logprob_tokens": [ + -1.1371183395385742 + ], + "logprob_reward": -0.3871183395385742, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871183395385742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08108791708946228, + 0.047019921243190765, + 0.08806674927473068, + 0.17104560136795044, + 0.043755266815423965, + 0.05806244909763336, + 0.12723198533058167, + 0.17079712450504303, + 0.2039770781993866, + 0.008955868892371655 + ], + "credit_map": [ + 0.009263238869607449, + 0.00819447822868824, + 0.010188213549554348, + 0.009042203426361084, + 0.009093211963772774, + 0.009608029387891293, + 0.009220433421432972, + 0.01647811383008957, + 0.003737666644155979, + 0.005278346594423056, + 0.004357724450528622, + 0.0073568448424339294, + 0.00702192448079586, + 0.005421609617769718, + 0.009317831136286259, + 0.004527977202087641, + 0.013123088516294956, + 0.011425786651670933, + 0.012743744067847729, + 0.014158173464238644, + 0.010401860810816288, + 0.00935454573482275, + 0.011070226319134235, + 0.005789328366518021, + 0.011458049528300762, + 0.00829707644879818, + 0.04849478229880333, + 0.009972678497433662, + 0.011333763599395752, + 0.05522226169705391, + 0.007446068339049816, + 0.018820911645889282, + 0.004888002295047045, + 0.005195552948862314, + 0.004375933203846216, + 0.008009080775082111, + 0.005293119233101606, + 0.0054762111976742744, + 0.005276929587125778, + 0.005240437109023333, + 0.011241345666348934, + 0.006864576600492001, + 0.008183183148503304, + 0.005543823353946209, + 0.005732796620577574, + 0.006870142184197903, + 0.00584412133321166, + 0.007782462518662214, + 0.025563843548297882, + 0.015278385020792484, + 0.018832599744200706, + 0.00935313943773508, + 0.011732686311006546, + 0.012269108556210995, + 0.011912617832422256, + 0.022289592772722244, + 0.021940935403108597, + 0.01080007292330265, + 0.042399339377880096, + 0.009818869642913342, + 0.01328697707504034, + 0.05140987038612366, + 0.00923377089202404, + 0.01190727949142456, + 0.03055090457201004, + 0.016515076160430908, + 0.03089177794754505, + 0.018825341016054153, + 0.016307953745126724, + 0.04507492110133171, + 0.016810256987810135, + 0.029000835493206978, + 0.000845597474835813, + 0.0007723596063442528, + 0.0011674334527924657, + 0.0010942543158307672, + 0.001206555520184338, + 0.0012481282465159893, + 0.0011574008967727423, + 0.0014641395537182689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.3871183395385742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029825976118445396 + ], + "opposite_logprob_tokens": [ + -3.5298259258270264 + ], + "logprob_reward": -0.029825976118445396, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029825976118445396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1258048117160797, + 0.04378848522901535, + 0.03787357732653618, + 0.29742398858070374, + 0.028095534071326256, + 0.0948631539940834, + 0.0707658901810646, + 0.09715808182954788, + 0.09903018176555634, + 0.10519628971815109 + ], + "credit_map": [ + 0.016286153346300125, + 0.015613782219588757, + 0.019621556624770164, + 0.01853393204510212, + 0.011470259167253971, + 0.012931804172694683, + 0.013320843689143658, + 0.018026478588581085, + 0.007721663452684879, + 0.005294165574014187, + 0.006414765026420355, + 0.004543054383248091, + 0.004033805802464485, + 0.004171115346252918, + 0.00643711956217885, + 0.005172797478735447, + 0.005137171596288681, + 0.004481869284063578, + 0.007063010707497597, + 0.004009447526186705, + 0.004181517753750086, + 0.004641817416995764, + 0.0034978645853698254, + 0.004860876593738794, + 0.02113323286175728, + 0.01543681975454092, + 0.08864454180002213, + 0.022396422922611237, + 0.0173978041857481, + 0.07409191876649857, + 0.024854905903339386, + 0.03346832096576691, + 0.0027805950958281755, + 0.0039013875648379326, + 0.002501554787158966, + 0.0033709655981510878, + 0.004526559729129076, + 0.0030755067709833384, + 0.003774904413148761, + 0.004164059180766344, + 0.0210161954164505, + 0.009029720909893513, + 0.014427903108298779, + 0.008481343276798725, + 0.00929456576704979, + 0.014606278389692307, + 0.006058690138161182, + 0.01194845698773861, + 0.016964994370937347, + 0.009304841980338097, + 0.01093309186398983, + 0.005720076151192188, + 0.004178989212960005, + 0.007200175896286964, + 0.006401740480214357, + 0.010061982087790966, + 0.009890496730804443, + 0.008011303842067719, + 0.021156253293156624, + 0.005285296589136124, + 0.007586506195366383, + 0.029146898537874222, + 0.006690992508083582, + 0.00939033180475235, + 0.01014623325318098, + 0.009632810950279236, + 0.013394549489021301, + 0.011300337500870228, + 0.009144552052021027, + 0.014959950000047684, + 0.018416564911603928, + 0.012035181745886803, + 0.01191119197756052, + 0.009291121736168861, + 0.012634743005037308, + 0.007949307560920715, + 0.013813166879117489, + 0.02058521844446659, + 0.00917093362659216, + 0.019840601831674576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.029825976118445396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948103964328766 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948103964328766, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948103964328766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04176569730043411, + 0.05961902067065239, + 0.031412385404109955, + 0.06234036013484001, + 0.028481323271989822, + 0.023674892261624336, + 0.030507368966937065, + 0.1847953498363495, + 0.17081822454929352, + 0.3665854036808014 + ], + "credit_map": [ + 0.005518055520951748, + 0.004939609672874212, + 0.0060073016211390495, + 0.007180873770266771, + 0.004485658369958401, + 0.00491271773353219, + 0.005680979695171118, + 0.0030405011493712664, + 0.010870452038943768, + 0.006200679112225771, + 0.00988057255744934, + 0.005664164200425148, + 0.006624198984354734, + 0.005246545188128948, + 0.007501170039176941, + 0.007631239015609026, + 0.004443590994924307, + 0.004265253897756338, + 0.004300102591514587, + 0.0040571922436356544, + 0.00397444935515523, + 0.0034380294382572174, + 0.003623535390943289, + 0.0033102333545684814, + 0.004659413360059261, + 0.004006354603916407, + 0.015510554425418377, + 0.005508293863385916, + 0.003548067295923829, + 0.016627071425318718, + 0.006810239981859922, + 0.0056703663431108, + 0.002474849810823798, + 0.0022557678166776896, + 0.002315888414159417, + 0.002222237642854452, + 0.003709697863087058, + 0.005689265672117472, + 0.0024163464549928904, + 0.007397268898785114, + 0.00376503961160779, + 0.002512935781851411, + 0.00384352495893836, + 0.002850737189874053, + 0.00268495362251997, + 0.002856424544006586, + 0.0017697252333164215, + 0.0033915522508323193, + 0.003912866581231356, + 0.0030522041488438845, + 0.0043371752835810184, + 0.0025084957014769316, + 0.003497554687783122, + 0.003924924414604902, + 0.0034073470160365105, + 0.0058668022975325584, + 0.021660273894667625, + 0.01375065091997385, + 0.04231305047869682, + 0.01203544344753027, + 0.011582748964428902, + 0.0491262786090374, + 0.017219655215740204, + 0.017107250168919563, + 0.024927301332354546, + 0.016708707436919212, + 0.01747051440179348, + 0.011516443453729153, + 0.0192339476197958, + 0.034151047468185425, + 0.024027060717344284, + 0.02278320863842964, + 0.03882426396012306, + 0.028358031064271927, + 0.05475619435310364, + 0.022715698927640915, + 0.03894395753741264, + 0.07005575299263, + 0.03868424519896507, + 0.07424724847078323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.005948103964328766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979542873799801 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.02979542873799801, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979542873799801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10810105502605438, + 0.07408053427934647, + 0.03635936602950096, + 0.23042051494121552, + 0.04685421288013458, + 0.14241068065166473, + 0.07345189154148102, + 0.16383770108222961, + 0.0633581355214119, + 0.06112586706876755 + ], + "credit_map": [ + 0.013640148565173149, + 0.013304688036441803, + 0.016425108537077904, + 0.015506826341152191, + 0.012734804302453995, + 0.014281042851507664, + 0.013835485093295574, + 0.008372961543500423, + 0.010700451210141182, + 0.006619919091463089, + 0.010969894006848335, + 0.010407446883618832, + 0.008857928216457367, + 0.006884359754621983, + 0.010319291613996029, + 0.009321250021457672, + 0.005146888084709644, + 0.003981531132012606, + 0.0045612044632434845, + 0.005498723592609167, + 0.00415761349722743, + 0.004443639423698187, + 0.005078510381281376, + 0.003491256618872285, + 0.014338018372654915, + 0.011592554859817028, + 0.06966182589530945, + 0.013900021091103554, + 0.013651175424456596, + 0.06504646688699722, + 0.01964886672794819, + 0.02258160710334778, + 0.00400296226143837, + 0.005658216308802366, + 0.0031619786750525236, + 0.006564210169017315, + 0.007050870917737484, + 0.0068657733500003815, + 0.005751990247517824, + 0.007798216305673122, + 0.01608182303607464, + 0.01672820746898651, + 0.02123301848769188, + 0.017550267279148102, + 0.01483144611120224, + 0.020498160272836685, + 0.016086598858237267, + 0.019401177763938904, + 0.0065366546623408794, + 0.010188570246100426, + 0.00862033199518919, + 0.006678330712020397, + 0.008475400507450104, + 0.007423854433000088, + 0.010552492924034595, + 0.014976262114942074, + 0.022641919553279877, + 0.01231813058257103, + 0.03843271732330322, + 0.010159040801227093, + 0.014165167696774006, + 0.04053077846765518, + 0.010030852630734444, + 0.015559096820652485, + 0.014983262866735458, + 0.005756756756454706, + 0.005674205720424652, + 0.004371764604002237, + 0.0050121331587433815, + 0.014927849173545837, + 0.004442171659320593, + 0.008190001361072063, + 0.010267231613397598, + 0.005236928816884756, + 0.005898731295019388, + 0.005594147834926844, + 0.0046654571779072285, + 0.012742813676595688, + 0.005418853834271431, + 0.011301705613732338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.02979542873799801 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, orbiting medium shot captures a female figure skater in poetic slow motion, executing a triple axel at the pinnacle of her performance. The camera moves fluidly around her, synchronized with her rotation, locking the viewer's gaze firmly on her. She wears a deep blue competition suit, studded with countless shimmering crystals. The Lycra material clings to her toned and graceful figure. As she spins at high speed, her dark brown hair, tightly gathered in a bun, escapes, tracing graceful arcs through the air. In slow motion, she is at the apex of her spin, her arms clasped tightly against her chest to maximize angular velocity. Her posture is impeccable, demonstrating supreme focus and control. The intense spotlight from the stadium's ceiling casts a bright, dramatic rim light across her figure, completely separating her from the background. This light not only emphasizes the contours of her muscles but also makes each crystal on her suit radiate a star-like brilliance. Behind her, the vast auditorium blurred into a soft, shimmering backdrop of varying shades of color, like a flowing Impressionist painting. The entire scene, reflected off the mirror-like surface of the ice, was imbued with a tension of power and beauty, creating a breathtaking, almost sacred atmosphere of competition.", + "seed": 16694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_14/index_4429_seed_16694_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": -0.15081624127924442, + "vlm_logprob": -0.15081624127924442, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.15081624127924442, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018578795716166496 + ], + "opposite_logprob_tokens": [ + -4.143579006195068 + ], + "logprob_reward": -0.018578795716166496, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018578795716166496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19173012673854828, + 0.1003330871462822, + 0.13045044243335724, + 0.025477109476923943, + 0.07181724905967712, + 0.12642937898635864, + 0.10937640070915222, + 0.15772384405136108, + 0.025411998853087425, + 0.06125035136938095 + ], + "credit_map": [ + 0.018089817836880684, + 0.01471757609397173, + 0.024135883897542953, + 0.021166255697607994, + 0.02508578449487686, + 0.02263333462178707, + 0.03234023228287697, + 0.03356124088168144, + 0.008218306116759777, + 0.012691482901573181, + 0.02352002076804638, + 0.01516078133136034, + 0.007028344087302685, + 0.011967038735747337, + 0.01078784093260765, + 0.01095927506685257, + 0.035929612815380096, + 0.0076917363330721855, + 0.03907233849167824, + 0.012707767076790333, + 0.009215230122208595, + 0.007957002148032188, + 0.010350327007472515, + 0.007526431232690811, + 0.0018395809456706047, + 0.0034748842008411884, + 0.0028162768576294184, + 0.0033226620871573687, + 0.0042566279880702496, + 0.0027805808931589127, + 0.004082769621163607, + 0.0029037268832325935, + 0.00890256930142641, + 0.009570351801812649, + 0.006048453971743584, + 0.005813463591039181, + 0.017095260322093964, + 0.009380974806845188, + 0.008520329371094704, + 0.0064858440309762955, + 0.01739460602402687, + 0.036344319581985474, + 0.02036454901099205, + 0.01000217441469431, + 0.011195519007742405, + 0.006272397469729185, + 0.012657921761274338, + 0.012197891250252724, + 0.012289086356759071, + 0.013459297828376293, + 0.01638302020728588, + 0.01277154590934515, + 0.017649373039603233, + 0.014691906049847603, + 0.01254881452769041, + 0.009583348408341408, + 0.017979448661208153, + 0.010177983902394772, + 0.055390067398548126, + 0.010967879556119442, + 0.008974917232990265, + 0.014907949604094028, + 0.00972814578562975, + 0.0295974463224411, + 0.001969024073332548, + 0.0022424538619816303, + 0.0031901434995234013, + 0.0044052754528820515, + 0.0030174122657626867, + 0.002290977630764246, + 0.004585465881973505, + 0.0037112447898834944, + 0.006033904850482941, + 0.007817716337740421, + 0.008517210371792316, + 0.006868554279208183, + 0.010799107141792774, + 0.006918853148818016, + 0.009484575130045414, + 0.004810428712517023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.018578795716166496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328283548355103 + ], + "opposite_logprob_tokens": [ + -0.7578283548355103 + ], + "logprob_reward": -0.6328283548355103, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328283548355103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08074267208576202, + 0.505425214767456, + 0.05257301405072212, + 0.04046168923377991, + 0.07119105756282806, + 0.04659244790673256, + 0.07502508908510208, + 0.04072146490216255, + 0.04666565731167793, + 0.04060163348913193 + ], + "credit_map": [ + 0.007337398361414671, + 0.007718669716268778, + 0.012860756367444992, + 0.010804539546370506, + 0.01049785129725933, + 0.010644515044987202, + 0.011124324053525925, + 0.009754628874361515, + 0.03168796747922897, + 0.061878502368927, + 0.1398283690214157, + 0.037065982818603516, + 0.05263286456465721, + 0.024794403463602066, + 0.08516904711723328, + 0.07236810773611069, + 0.011043436825275421, + 0.004057436250150204, + 0.009954376146197319, + 0.005123530048877001, + 0.0066719683818519115, + 0.004738416988402605, + 0.006533284205943346, + 0.0044505721889436245, + 0.00321528734639287, + 0.009843189269304276, + 0.00821533054113388, + 0.003115121042355895, + 0.0056158448569476604, + 0.0027371433097869158, + 0.0037653781473636627, + 0.003954402636736631, + 0.006577277556061745, + 0.007604824844747782, + 0.005453544668853283, + 0.007633404340595007, + 0.016720736399292946, + 0.010896292515099049, + 0.010985277593135834, + 0.005319702904671431, + 0.005874915514141321, + 0.00814327597618103, + 0.005697771906852722, + 0.00729792844504118, + 0.006474468391388655, + 0.002343327971175313, + 0.006073660217225552, + 0.004687104374170303, + 0.005790363997220993, + 0.009024243801832199, + 0.01140274666249752, + 0.008393757045269012, + 0.011344020254909992, + 0.009439817629754543, + 0.011499617248773575, + 0.008130536414682865, + 0.006591026205569506, + 0.0039238957688212395, + 0.010149635374546051, + 0.003085118718445301, + 0.0037893974222242832, + 0.005317434202879667, + 0.004174940288066864, + 0.0036900201812386513, + 0.0032783220522105694, + 0.004896773491054773, + 0.005666267592459917, + 0.008232067339122295, + 0.007411154452711344, + 0.003934288863092661, + 0.00785047933459282, + 0.0053963069804012775, + 0.005236664321273565, + 0.004949002992361784, + 0.006940076127648354, + 0.004428638610988855, + 0.004220848437398672, + 0.0056083728559315205, + 0.005455576814711094, + 0.0037624556571245193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.6328283548355103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498092621564865 + ], + "opposite_logprob_tokens": [ + -2.929980993270874 + ], + "logprob_reward": -0.05498092621564865, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498092621564865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.100980244576931, + 0.17708563804626465, + 0.067508764564991, + 0.07560867071151733, + 0.21446731686592102, + 0.08167067915201187, + 0.08900707960128784, + 0.08766746520996094, + 0.047837287187576294, + 0.05816691368818283 + ], + "credit_map": [ + 0.013638511300086975, + 0.011276353150606155, + 0.017395764589309692, + 0.012050165794789791, + 0.00796435121446848, + 0.01279628649353981, + 0.011477564461529255, + 0.014381240122020245, + 0.010304661467671394, + 0.030786551535129547, + 0.06101309880614281, + 0.026462292298674583, + 0.00952074397355318, + 0.012040775269269943, + 0.010790937580168247, + 0.016166554763913155, + 0.015962660312652588, + 0.005617820657789707, + 0.016665015369653702, + 0.006763478275388479, + 0.0047469958662986755, + 0.005474801175296307, + 0.005901585333049297, + 0.0063764010556042194, + 0.006091515999287367, + 0.023447532206773758, + 0.02051859349012375, + 0.006031682249158621, + 0.005849123932421207, + 0.0051429495215415955, + 0.0035228622145950794, + 0.005004399921745062, + 0.023777347058057785, + 0.0250154510140419, + 0.017785029485821724, + 0.019721616059541702, + 0.044349830597639084, + 0.026629146188497543, + 0.037788353860378265, + 0.01940053328871727, + 0.009135155007243156, + 0.013356635347008705, + 0.012641499750316143, + 0.00826894398778677, + 0.01425285916775465, + 0.004575591534376144, + 0.014278233982622623, + 0.005161753855645657, + 0.007294422015547752, + 0.017329389229416847, + 0.013698285445570946, + 0.010175301693379879, + 0.011311455629765987, + 0.010886196047067642, + 0.011593891307711601, + 0.006718123331665993, + 0.013417651876807213, + 0.006998030003160238, + 0.010157512500882149, + 0.008122353814542294, + 0.006512586027383804, + 0.01046773511916399, + 0.009757532738149166, + 0.022234058007597923, + 0.004092549905180931, + 0.004763311240822077, + 0.006919786334037781, + 0.008569732308387756, + 0.005315816029906273, + 0.0038109971210360527, + 0.008145815692842007, + 0.006219273898750544, + 0.006703751627355814, + 0.006667886860668659, + 0.01025538519024849, + 0.006358986720442772, + 0.006207752972841263, + 0.009506653994321823, + 0.006523689720779657, + 0.005942800082266331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.05498092621564865 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635896533727646 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635896533727646, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635896533727646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11859481781721115, + 0.40458521246910095, + 0.04287637397646904, + 0.039472680538892746, + 0.047887057065963745, + 0.11802539229393005, + 0.08156296610832214, + 0.05508532375097275, + 0.022442445158958435, + 0.06946767121553421 + ], + "credit_map": [ + 0.01107990462332964, + 0.008145051077008247, + 0.025005090981721878, + 0.01876264251768589, + 0.010796231217682362, + 0.01525840163230896, + 0.011710410006344318, + 0.01783708855509758, + 0.02340962551534176, + 0.04603443667292595, + 0.15586790442466736, + 0.0513167679309845, + 0.018281864002346992, + 0.02131696417927742, + 0.018477702513337135, + 0.06987998634576797, + 0.0060566789470613, + 0.00811756681650877, + 0.004591853823512793, + 0.007480539381504059, + 0.0043392847292125225, + 0.0041814097203314304, + 0.003393243532627821, + 0.004715796560049057, + 0.0030398936942219734, + 0.006340217776596546, + 0.006744552403688431, + 0.0055868616327643394, + 0.0044451067224144936, + 0.003919902723282576, + 0.004649985581636429, + 0.004746158607304096, + 0.005662319716066122, + 0.005150858778506517, + 0.004239767324179411, + 0.003736727172508836, + 0.010808776132762432, + 0.005763825960457325, + 0.008338628336787224, + 0.004186155274510384, + 0.016733946278691292, + 0.027624433860182762, + 0.024982737377285957, + 0.012498104944825172, + 0.011952743865549564, + 0.005086931400001049, + 0.009914038702845573, + 0.009232458658516407, + 0.0073878709226846695, + 0.013385803438723087, + 0.014143376611173153, + 0.00789803359657526, + 0.012231962755322456, + 0.010212705470621586, + 0.008673237636685371, + 0.007629971485584974, + 0.006237698253244162, + 0.004178131930530071, + 0.014283521100878716, + 0.005627451464533806, + 0.005264149978756905, + 0.005530346650630236, + 0.005769102368503809, + 0.008194921538233757, + 0.0018247406696900725, + 0.0017189907375723124, + 0.003440078115090728, + 0.004273310303688049, + 0.0023729733657091856, + 0.0029011722654104233, + 0.0030366259161382914, + 0.002874554367735982, + 0.007615027017891407, + 0.007691378239542246, + 0.011278285644948483, + 0.006702091079205275, + 0.007950733415782452, + 0.00852529238909483, + 0.009808561764657497, + 0.009896297007799149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.004635896533727646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305723309516907 + ], + "opposite_logprob_tokens": [ + -3.1680572032928467 + ], + "logprob_reward": -0.04305723309516907, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305723309516907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08486008644104004, + 0.1971147507429123, + 0.07510345429182053, + 0.10054942965507507, + 0.1269690990447998, + 0.10300029814243317, + 0.08654472231864929, + 0.160429909825325, + 0.030077911913394928, + 0.03535032272338867 + ], + "credit_map": [ + 0.0070016346871852875, + 0.00648849131539464, + 0.011840333230793476, + 0.009386617690324783, + 0.012654440477490425, + 0.011389179155230522, + 0.013964174315333366, + 0.012135215103626251, + 0.01643892377614975, + 0.026554381474852562, + 0.048783447593450546, + 0.021492451429367065, + 0.016917094588279724, + 0.019462918862700462, + 0.016350042074918747, + 0.03111548349261284, + 0.0192874763160944, + 0.006019131746143103, + 0.018311988562345505, + 0.008373740129172802, + 0.005916284862905741, + 0.006128175184130669, + 0.005314819514751434, + 0.0057518393732607365, + 0.006630634423345327, + 0.02435954473912716, + 0.0242062546312809, + 0.010733946226537228, + 0.010617805644869804, + 0.007898068986833096, + 0.008937477134168148, + 0.007165696006268263, + 0.012479588389396667, + 0.01280766911804676, + 0.010555257089436054, + 0.01079231221228838, + 0.022004203870892525, + 0.024613307788968086, + 0.022783886641263962, + 0.010932871140539646, + 0.012628813274204731, + 0.026109665632247925, + 0.02100544609129429, + 0.008205077610909939, + 0.01166487019509077, + 0.005389613099396229, + 0.01001677569001913, + 0.007980029098689556, + 0.009231268428266048, + 0.01316518522799015, + 0.013802586123347282, + 0.006557205226272345, + 0.010169360786676407, + 0.011318635195493698, + 0.011340691708028316, + 0.010959790088236332, + 0.018871288746595383, + 0.01089455559849739, + 0.05758790671825409, + 0.010750895366072655, + 0.007950296625494957, + 0.018323617056012154, + 0.01001217495650053, + 0.026039177551865578, + 0.0024403438437730074, + 0.0026644712779670954, + 0.003921017050743103, + 0.003911343868821859, + 0.004406817257404327, + 0.0030500644352287054, + 0.005250693764537573, + 0.004433161113411188, + 0.004020509775727987, + 0.004762014839798212, + 0.005946415942162275, + 0.0033264539670199156, + 0.004961858969181776, + 0.005560512654483318, + 0.0037174448370933533, + 0.0030551140662282705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.04305723309516907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, orbiting medium shot captures a female figure skater in poetic slow motion, executing a triple axel at the pinnacle of her performance. The camera moves fluidly around her, synchronized with her rotation, locking the viewer's gaze firmly on her. She wears a deep blue competition suit, studded with countless shimmering crystals. The Lycra material clings to her toned and graceful figure. As she spins at high speed, her dark brown hair, tightly gathered in a bun, escapes, tracing graceful arcs through the air. In slow motion, she is at the apex of her spin, her arms clasped tightly against her chest to maximize angular velocity. Her posture is impeccable, demonstrating supreme focus and control. The intense spotlight from the stadium's ceiling casts a bright, dramatic rim light across her figure, completely separating her from the background. This light not only emphasizes the contours of her muscles but also makes each crystal on her suit radiate a star-like brilliance. Behind her, the vast auditorium blurred into a soft, shimmering backdrop of varying shades of color, like a flowing Impressionist painting. The entire scene, reflected off the mirror-like surface of the ice, was imbued with a tension of power and beauty, creating a breathtaking, almost sacred atmosphere of competition.", + "seed": 16695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_14/index_4429_seed_16695_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": -0.10738301435485482, + "vlm_logprob": -0.10738301435485482, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.10738301435485482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013585491105914116 + ], + "opposite_logprob_tokens": [ + -4.513585567474365 + ], + "logprob_reward": -0.013585491105914116, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013585491105914116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07099894434213638, + 0.02464345283806324, + 0.04052326828241348, + 0.10653387755155563, + 0.0728984996676445, + 0.4949522912502289, + 0.02889089845120907, + 0.029499560594558716, + 0.04366467520594597, + 0.0873946025967598 + ], + "credit_map": [ + 0.006037715822458267, + 0.0088252704590559, + 0.007717125117778778, + 0.007886774837970734, + 0.007207568269222975, + 0.009352907538414001, + 0.013290232978761196, + 0.010681338608264923, + 0.002360088750720024, + 0.0040343585424125195, + 0.002445494057610631, + 0.0044884998351335526, + 0.0026521540712565184, + 0.0019241892732679844, + 0.0033525293692946434, + 0.0033861356787383556, + 0.007004456594586372, + 0.004174599424004555, + 0.008582684211432934, + 0.0033768839202821255, + 0.0029021131340414286, + 0.00638218317180872, + 0.0030520993750542402, + 0.005048243794590235, + 0.009910428896546364, + 0.007769859861582518, + 0.03105604276061058, + 0.008001066744327545, + 0.006973245646804571, + 0.012514148838818073, + 0.010243880562484264, + 0.020065193995833397, + 0.007582295220345259, + 0.005641614552587271, + 0.016467487439513206, + 0.0075074308551847935, + 0.007319174241274595, + 0.011624432168900967, + 0.0065553090535104275, + 0.010200746357440948, + 0.0960645079612732, + 0.04729278385639191, + 0.07863455265760422, + 0.046529922634363174, + 0.05109954625368118, + 0.05525645613670349, + 0.07360851764678955, + 0.04646594822406769, + 0.0030297776684165, + 0.0017759238835424185, + 0.00515908794477582, + 0.0031324089504778385, + 0.002824309514835477, + 0.006991592701524496, + 0.0024314746260643005, + 0.003546320367604494, + 0.0036111341323703527, + 0.001882556127384305, + 0.005098295398056507, + 0.003101153764873743, + 0.002809694269672036, + 0.005337219685316086, + 0.0031605602707713842, + 0.004498941823840141, + 0.003785148961469531, + 0.0034712052438408136, + 0.008462268859148026, + 0.004904260393232107, + 0.004647535737603903, + 0.003947983495891094, + 0.004769050050526857, + 0.009677217341959476, + 0.008303491398692131, + 0.005714148283004761, + 0.016936516389250755, + 0.008394770324230194, + 0.00889087375253439, + 0.010836286470293999, + 0.008672159165143967, + 0.01964634284377098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.013585491105914116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.428961843252182 + ], + "opposite_logprob_tokens": [ + -1.0539618730545044 + ], + "logprob_reward": -0.428961843252182, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.428961843252182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14949733018875122, + 0.07053082436323166, + 0.13168694078922272, + 0.0891120657324791, + 0.08760738372802734, + 0.05206013843417168, + 0.22390146553516388, + 0.10082411020994186, + 0.05131238326430321, + 0.04346732422709465 + ], + "credit_map": [ + 0.01540432870388031, + 0.019126402214169502, + 0.022430473938584328, + 0.018813831731677055, + 0.017955584451556206, + 0.017349690198898315, + 0.025750253349542618, + 0.012666766531765461, + 0.00832635723054409, + 0.012101716361939907, + 0.009891632944345474, + 0.012261615134775639, + 0.008091786876320839, + 0.004150812979787588, + 0.007690158672630787, + 0.008016747422516346, + 0.013212818652391434, + 0.014135614968836308, + 0.02344345673918724, + 0.0117368558421731, + 0.013472317717969418, + 0.026043199002742767, + 0.014696410857141018, + 0.014946273528039455, + 0.025976242497563362, + 0.007107069715857506, + 0.019762352108955383, + 0.004315676633268595, + 0.003433960722759366, + 0.011726644821465015, + 0.0036775388289242983, + 0.013112585060298443, + 0.007187835406512022, + 0.006529377773404121, + 0.018452461808919907, + 0.006378662772476673, + 0.013053048402071, + 0.009111951105296612, + 0.013771316036581993, + 0.013122732751071453, + 0.005882568657398224, + 0.004649856127798557, + 0.006564721930772066, + 0.00604200828820467, + 0.007196175865828991, + 0.007674055173993111, + 0.00790442992001772, + 0.00614631874486804, + 0.029643278568983078, + 0.01876228302717209, + 0.03374788165092468, + 0.026278747245669365, + 0.026903342455625534, + 0.034625500440597534, + 0.02341032400727272, + 0.03053010068833828, + 0.009321320801973343, + 0.008389883674681187, + 0.01823609136044979, + 0.012880878522992134, + 0.013093848712742329, + 0.013065841048955917, + 0.01187929604202509, + 0.013956954702734947, + 0.007559089921414852, + 0.005509063135832548, + 0.00827693473547697, + 0.005604066886007786, + 0.008806372992694378, + 0.004926091525703669, + 0.0047362265177071095, + 0.0058945356868207455, + 0.004163859412074089, + 0.0027276023756712675, + 0.009994924999773502, + 0.00310736196115613, + 0.0033887557219713926, + 0.005404692143201828, + 0.00386831839568913, + 0.01081180665642023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.428961843252182 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207505986094475 + ], + "opposite_logprob_tokens": [ + -2.812075138092041 + ], + "logprob_reward": -0.06207505986094475, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207505986094475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1004490852355957, + 0.07956646382808685, + 0.08007381111383438, + 0.1030253991484642, + 0.05885900557041168, + 0.07527868449687958, + 0.20853972434997559, + 0.08092948794364929, + 0.06644129753112793, + 0.146837055683136 + ], + "credit_map": [ + 0.010578246787190437, + 0.010452298447489738, + 0.009875313378870487, + 0.012439711950719357, + 0.012356186285614967, + 0.011889765970408916, + 0.022016307339072227, + 0.010841256938874722, + 0.0070798457600176334, + 0.009234708733856678, + 0.01019960641860962, + 0.017331914976239204, + 0.0075669861398637295, + 0.006628113798797131, + 0.011499990709125996, + 0.010025295428931713, + 0.010630827397108078, + 0.008107732981443405, + 0.01904069259762764, + 0.007769468706101179, + 0.004838345572352409, + 0.012422683648765087, + 0.007246148306876421, + 0.010017916560173035, + 0.014668759889900684, + 0.0066308025270700455, + 0.03070814535021782, + 0.006567425560206175, + 0.005421040114015341, + 0.011542657390236855, + 0.005911818705499172, + 0.02157473936676979, + 0.006944320164620876, + 0.005616109352558851, + 0.009352175518870354, + 0.005303855054080486, + 0.0077700018882751465, + 0.009811228141188622, + 0.006549046374857426, + 0.007512277457863092, + 0.007330257911235094, + 0.006667743436992168, + 0.015032602474093437, + 0.005985647439956665, + 0.007434830069541931, + 0.013348491862416267, + 0.00912562757730484, + 0.010353482328355312, + 0.024723073467612267, + 0.015122875571250916, + 0.034148670732975006, + 0.02416262775659561, + 0.023835891857743263, + 0.03814271464943886, + 0.021552512422204018, + 0.026851341128349304, + 0.006510587874799967, + 0.008072759956121445, + 0.010280213318765163, + 0.011034093797206879, + 0.009400639683008194, + 0.013957766816020012, + 0.011175866238772869, + 0.01049755234271288, + 0.01071578823029995, + 0.006994405761361122, + 0.009391902945935726, + 0.009842564351856709, + 0.00675916438922286, + 0.00831247866153717, + 0.005320358090102673, + 0.009104632772505283, + 0.01814325340092182, + 0.009610955603420734, + 0.030308635905385017, + 0.011344821192324162, + 0.01250623818486929, + 0.013570785522460938, + 0.01338845957070589, + 0.03796390816569328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.06207505986094475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18564602732658386, + 0.06691223382949829, + 0.05706241354346275, + 0.056267425417900085, + 0.17087654769420624, + 0.1059771329164505, + 0.041723571717739105, + 0.06824204325675964, + 0.13993017375469208, + 0.10736244171857834 + ], + "credit_map": [ + 0.024234281852841377, + 0.01605224609375, + 0.023935431614518166, + 0.018752261996269226, + 0.018167808651924133, + 0.024366721510887146, + 0.028543522581458092, + 0.031593747437000275, + 0.007669673301279545, + 0.005844329949468374, + 0.012583368457853794, + 0.010202796198427677, + 0.007959840819239616, + 0.00559097807854414, + 0.008252019993960857, + 0.008809227496385574, + 0.011021646670997143, + 0.0064730108715593815, + 0.01071198657155037, + 0.00449984148144722, + 0.004532285500317812, + 0.0084248511120677, + 0.0032069680746644735, + 0.008191823028028011, + 0.008027777075767517, + 0.0053285700269043446, + 0.015421015210449696, + 0.004153397399932146, + 0.003216842655092478, + 0.006801753304898739, + 0.0031571509316563606, + 0.010160918347537518, + 0.023617183789610863, + 0.015544992871582508, + 0.03821638971567154, + 0.015022577717900276, + 0.012952703051269054, + 0.03518734872341156, + 0.00916928332298994, + 0.02116606943309307, + 0.013739208690822124, + 0.009827674366533756, + 0.021785704419016838, + 0.00694095017388463, + 0.00803349819034338, + 0.020373692736029625, + 0.008311189711093903, + 0.016965214163064957, + 0.0066379401832818985, + 0.0030797605868428946, + 0.006839019246399403, + 0.003041512332856655, + 0.003414770355448127, + 0.0072408998385071754, + 0.004089963156729937, + 0.007379706017673016, + 0.008753180503845215, + 0.004840866196900606, + 0.010461458936333656, + 0.006069875787943602, + 0.007240526843816042, + 0.010467978194355965, + 0.009300106205046177, + 0.011108052916824818, + 0.02904789336025715, + 0.006504345219582319, + 0.023220786824822426, + 0.008893837220966816, + 0.01073224563151598, + 0.018183844164013863, + 0.009354067966341972, + 0.033993154764175415, + 0.015681087970733643, + 0.006265809293836355, + 0.022552739828824997, + 0.0067560868337750435, + 0.006572199519723654, + 0.016148483380675316, + 0.009278521873056889, + 0.024107512086629868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female figure skater wearing a deep blue competition suit studded with crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skater executing a triple axel spin with her arms clasped tightly against her chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skater's dark brown hair tied in a bun that is escaping and tracing arcs through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a blurred auditorium with varying shades of color resembling an Impressionist painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a steady, orbiting medium shot that moves fluidly around the skater?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344336569309235 + ], + "opposite_logprob_tokens": [ + -3.6513442993164062 + ], + "logprob_reward": -0.026344336569309235, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344336569309235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2548704743385315, + 0.09520809352397919, + 0.046673789620399475, + 0.1627950221300125, + 0.0435800738632679, + 0.05638686940073967, + 0.056434184312820435, + 0.09036565572023392, + 0.058070454746484756, + 0.13561539351940155 + ], + "credit_map": [ + 0.03892803192138672, + 0.024277593940496445, + 0.03367455676198006, + 0.022405186668038368, + 0.021345170214772224, + 0.028706135228276253, + 0.03674554079771042, + 0.048788223415613174, + 0.01022768672555685, + 0.010559272952377796, + 0.014047637581825256, + 0.01727885939180851, + 0.012166243977844715, + 0.007059692870825529, + 0.01062958873808384, + 0.013239107094705105, + 0.005631018429994583, + 0.007275134790688753, + 0.008149080909788609, + 0.004736737813800573, + 0.003033369779586792, + 0.00809274148195982, + 0.004604119807481766, + 0.005151584278792143, + 0.034401215612888336, + 0.009538333863019943, + 0.04263206198811531, + 0.009668415412306786, + 0.008243663236498833, + 0.018785415217280388, + 0.009050215594470501, + 0.030475681647658348, + 0.004333270248025656, + 0.0039171199314296246, + 0.006807180121541023, + 0.0028721385169774294, + 0.007572029251605272, + 0.007776746992021799, + 0.00414208322763443, + 0.00615950021892786, + 0.008776634931564331, + 0.00519508495926857, + 0.008891099132597446, + 0.0041808756068348885, + 0.004953386727720499, + 0.009436903521418571, + 0.008609657175838947, + 0.006343221291899681, + 0.005761474836617708, + 0.0030357306823134422, + 0.011027449741959572, + 0.005201089195907116, + 0.004988323897123337, + 0.013375299051404, + 0.005344081204384565, + 0.007700727321207523, + 0.011900286190211773, + 0.0068781389854848385, + 0.01803642511367798, + 0.007994869723916054, + 0.008412103168666363, + 0.013558680191636086, + 0.00766709353774786, + 0.01591804251074791, + 0.006664056796580553, + 0.005729063879698515, + 0.009254117496311665, + 0.006594163365662098, + 0.008182798512279987, + 0.00530937360599637, + 0.007087215315550566, + 0.009249663911759853, + 0.014151031151413918, + 0.006307866889983416, + 0.034499797970056534, + 0.00912547018378973, + 0.008129940368235111, + 0.013848146423697472, + 0.009473387151956558, + 0.0400797463953495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.026344336569309235 + } + ], + "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_15/index_tensor([4307, 4307])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_15/index_tensor([4307, 4307])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e4fc8b96044ae3dc1d4b30202ad51403d8c8bcb1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_15/index_tensor([4307, 4307])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide, long shot, captured at a low angle, pans backward and slowly rises as a train advances, ultimately capturing an epic drone-like tracking perspective. In the frame, a dark, weathered steam locomotive, towing several old, rusted iron carriages, jolted violently across a vast, barren plain. Two rusted tracks, with rotting sleepers, stretch across the cracked, earthy earth until they disappear into the horizon. The locomotive's towering smokestack spews erratically, emitting thick, gray-black steam and soot, mixed with dust, leaving a long, dissipating trail behind it. The train heaves and pitches violently on the uneven track, its joints repeatedly stretched and squeezed. Each jolt threatens to derail, a gesture of its tenacious struggle to move forward. The low-hanging sun shines through the gray-yellow clouds in the sky, casting dim and weak light, casting long shadows on the dry weeds and scattered bushes, dyeing the entire scene with a monotonous and depressing warm tone, creating a desolate, lonely and apocalyptic atmosphere full of primitive mechanical power.", + "seed": 20584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_15/index_4307_seed_20584_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": -0.019299288140609862, + "vlm_logprob": -0.019299288140609862, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.019299288140609862, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.018831750378012657 + ], + "opposite_logprob_tokens": [ + -4.268831729888916 + ], + "logprob_reward": -0.018831750378012657, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018831750378012657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02644089236855507, + 0.009291628375649452, + 0.013170747086405754, + 0.014058174565434456, + 0.011510268785059452, + 0.015967460349202156, + 0.8708797097206116, + 0.008722725324332714, + 0.011270791292190552, + 0.018687620759010315 + ], + "credit_map": [ + 0.002542427508160472, + 0.0044329240918159485, + 0.0025662649422883987, + 0.003495418932288885, + 0.0025277489330619574, + 0.0026515170466154814, + 0.003316074376925826, + 0.0049085165373981, + 0.0009013833478093147, + 0.0006936457357369363, + 0.001473484211601317, + 0.0008475107606500387, + 0.0009432879160158336, + 0.001024706056341529, + 0.001097147585824132, + 0.002310462761670351, + 0.001491603790782392, + 0.0011633149115368724, + 0.001892715459689498, + 0.001596210990101099, + 0.0010145003907382488, + 0.0018596429144963622, + 0.0019020203035324812, + 0.0022507384419441223, + 0.0013911151327192783, + 0.0014421044616028666, + 0.0011642391327768564, + 0.0011982530122622848, + 0.002249228535220027, + 0.0011915722861886024, + 0.0029078072402626276, + 0.0025138542987406254, + 0.000916461693122983, + 0.0008557841065339744, + 0.002542739501222968, + 0.0013522692024707794, + 0.0010923774680122733, + 0.0014956624945625663, + 0.001398439984768629, + 0.0018565350910648704, + 0.0013608450535684824, + 0.0009085166966542602, + 0.0022304989397525787, + 0.0016523041995242238, + 0.0012112656841054559, + 0.0012885899050161242, + 0.0020841287914663553, + 0.00523131201043725, + 0.0328829400241375, + 0.017278872430324554, + 0.36264005303382874, + 0.34335124492645264, + 0.017733674496412277, + 0.015423511154949665, + 0.02883429080247879, + 0.05273514613509178, + 0.0009005336905829608, + 0.0009616360766813159, + 0.0012108569499105215, + 0.0012107008369639516, + 0.0012194848386570811, + 0.0009081290336325765, + 0.0012901108711957932, + 0.0010212725028395653, + 0.0007953899330459535, + 0.0008305321680381894, + 0.0018623361829668283, + 0.0032258378341794014, + 0.0011178505374118686, + 0.0009396658861078322, + 0.0012718206271529198, + 0.0012273581232875586, + 0.0011633076937869191, + 0.0013058637268841267, + 0.0017207689816132188, + 0.0027978133875876665, + 0.0023173028603196144, + 0.0017201851587742567, + 0.004279484041035175, + 0.0033828949090093374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.018831750378012657 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.023294363170862198 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.023294363170862198, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294363170862198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17873995006084442, + 0.12404938787221909, + 0.045524246990680695, + 0.14021135866641998, + 0.06232744827866554, + 0.20433811843395233, + 0.1017509251832962, + 0.03246184438467026, + 0.026939520612359047, + 0.08365728706121445 + ], + "credit_map": [ + 0.015637747943401337, + 0.024040160700678825, + 0.022419223561882973, + 0.026249537244439125, + 0.014446360059082508, + 0.015491459518671036, + 0.021625731140375137, + 0.038829706609249115, + 0.015259585343301296, + 0.008259494788944721, + 0.01722475327551365, + 0.013161120936274529, + 0.009081043303012848, + 0.012329629622399807, + 0.01693466491997242, + 0.031799089163541794, + 0.004697776865214109, + 0.0034373824018985033, + 0.006275171879678965, + 0.0048585631884634495, + 0.004147634841501713, + 0.0059790643863379955, + 0.0052904896438121796, + 0.010838157497346401, + 0.011250600218772888, + 0.010670836083590984, + 0.01930764690041542, + 0.021335242316126823, + 0.010625975206494331, + 0.014329086989164352, + 0.029973015189170837, + 0.02271893434226513, + 0.005419556517153978, + 0.003439293010160327, + 0.0156813133507967, + 0.009229441173374653, + 0.004427189938724041, + 0.006956971250474453, + 0.009859854355454445, + 0.007313823327422142, + 0.01933923549950123, + 0.006289491429924965, + 0.03191617503762245, + 0.025556882843375206, + 0.006108756177127361, + 0.011623945087194443, + 0.020526399835944176, + 0.08297722786664963, + 0.006504382472485304, + 0.007286663167178631, + 0.025546161457896233, + 0.019968807697296143, + 0.011908132582902908, + 0.009833711199462414, + 0.010734155774116516, + 0.009968899190425873, + 0.0037679513916373253, + 0.003167098155245185, + 0.006586146075278521, + 0.004958288744091988, + 0.004110605921596289, + 0.004031826741993427, + 0.0030563771724700928, + 0.0027835462242364883, + 0.002223710063844919, + 0.002432438777759671, + 0.004772070795297623, + 0.004904486704617739, + 0.0029622181318700314, + 0.0021380565594881773, + 0.0035268485080450773, + 0.003979687578976154, + 0.0064981053583323956, + 0.005476599559187889, + 0.00870286114513874, + 0.015514607541263103, + 0.0050472235307097435, + 0.005901718977838755, + 0.023377517238259315, + 0.013138649053871632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.023294363170862198 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.033701326698064804 + ], + "opposite_logprob_tokens": [ + -3.4087014198303223 + ], + "logprob_reward": -0.033701326698064804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033701326698064804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3420026898384094, + 0.025440217927098274, + 0.037702810019254684, + 0.03215264901518822, + 0.026853585615754128, + 0.09283388406038284, + 0.3438873589038849, + 0.056143619120121, + 0.011687807738780975, + 0.03129534795880318 + ], + "credit_map": [ + 0.03104868344962597, + 0.044510647654533386, + 0.03788372874259949, + 0.049819767475128174, + 0.029596837237477303, + 0.030082007870078087, + 0.04096396267414093, + 0.07809708267450333, + 0.003181920852512121, + 0.0023956377990543842, + 0.0028397981077432632, + 0.0032335566356778145, + 0.001986660063266754, + 0.00264625228010118, + 0.003879463765770197, + 0.005276930518448353, + 0.004506924655288458, + 0.004059367347508669, + 0.005612989887595177, + 0.0033587845973670483, + 0.0047319866716861725, + 0.006112504284828901, + 0.0041190460324287415, + 0.005201209336519241, + 0.003491350682452321, + 0.0030473177321255207, + 0.004574132151901722, + 0.003918041475117207, + 0.0053378003649413586, + 0.00484892912209034, + 0.0037630705628544092, + 0.0031720090191811323, + 0.0022117372136563063, + 0.0013219205429777503, + 0.004655805416405201, + 0.002366521628573537, + 0.002063260180875659, + 0.0032394728623330593, + 0.004282752051949501, + 0.006712118163704872, + 0.008521169424057007, + 0.0030421235132962465, + 0.013730715028941631, + 0.010987043380737305, + 0.004446545150130987, + 0.005808087065815926, + 0.009209126234054565, + 0.03708907961845398, + 0.012543140910565853, + 0.007188071962445974, + 0.14137616753578186, + 0.1395067423582077, + 0.007049749605357647, + 0.006198609713464975, + 0.01196877472102642, + 0.018056122586131096, + 0.0050147599540650845, + 0.005043358542025089, + 0.011119263246655464, + 0.01553256157785654, + 0.004394939634948969, + 0.003780304454267025, + 0.005592738278210163, + 0.005665698554366827, + 0.0008617846760898829, + 0.0009090215899050236, + 0.002107937354594469, + 0.003100910922512412, + 0.001120837638154626, + 0.00098368467297405, + 0.001238395692780614, + 0.001365236472338438, + 0.002588895382359624, + 0.0023446993436664343, + 0.0032369906548410654, + 0.005548829212784767, + 0.0028490645345300436, + 0.0025114763993769884, + 0.006710374262183905, + 0.0055050235241651535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.033701326698064804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.018181100487709045 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.018181100487709045, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181100487709045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04559734836220741, + 0.03745397552847862, + 0.05311072990298271, + 0.06036347895860672, + 0.025513604283332825, + 0.02780267968773842, + 0.5117740035057068, + 0.061124954372644424, + 0.058037206530570984, + 0.11922204494476318 + ], + "credit_map": [ + 0.005270001012831926, + 0.005611254367977381, + 0.0037279194220900536, + 0.00633225217461586, + 0.0069703455083072186, + 0.00484920758754015, + 0.005486910231411457, + 0.007349458057433367, + 0.0024905793834477663, + 0.004069596529006958, + 0.004733921494334936, + 0.005106759257614613, + 0.006158835254609585, + 0.0033979415893554688, + 0.005178735591471195, + 0.006317604798823595, + 0.006478636991232634, + 0.006427736021578312, + 0.007977291941642761, + 0.0071147228591144085, + 0.005460172425955534, + 0.008112247101962566, + 0.00676894374191761, + 0.004770979285240173, + 0.006126943044364452, + 0.006577106658369303, + 0.0061652930453419685, + 0.007282483857125044, + 0.009436411783099174, + 0.006009681615978479, + 0.01053355261683464, + 0.00823200773447752, + 0.002511107362806797, + 0.0015753153711557388, + 0.006234883796423674, + 0.0025356230325996876, + 0.001673177583143115, + 0.0024481064174324274, + 0.0039273276925086975, + 0.0046080625616014, + 0.002032434567809105, + 0.0020509115420281887, + 0.003588668769225478, + 0.0036104819737374783, + 0.0024856780655682087, + 0.002890649251639843, + 0.0032697017304599285, + 0.007874155417084694, + 0.018971849232912064, + 0.012739366851747036, + 0.20511072874069214, + 0.19988979399204254, + 0.016506023705005646, + 0.012352501042187214, + 0.018150076270103455, + 0.028053661808371544, + 0.004298586398363113, + 0.0055217077024281025, + 0.012742012739181519, + 0.01144222728908062, + 0.008022968657314777, + 0.00530305877327919, + 0.006287039257586002, + 0.007507354021072388, + 0.003828787710517645, + 0.004105901811271906, + 0.010502426885068417, + 0.016605239361524582, + 0.005623801611363888, + 0.0046099950559437275, + 0.006689291447401047, + 0.006071761716157198, + 0.008309539407491684, + 0.00693500554189086, + 0.011895110830664635, + 0.02153969369828701, + 0.007926017045974731, + 0.010184546932578087, + 0.03184198588132858, + 0.020590154454112053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.018181100487709045 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.0024878999684005976 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024878999684005976, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024878999684005976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14079712331295013, + 0.02508842572569847, + 0.00945423636585474, + 0.01771126314997673, + 0.01040681917220354, + 0.0781768262386322, + 0.6459372043609619, + 0.029200782999396324, + 0.011906992644071579, + 0.031320374459028244 + ], + "credit_map": [ + 0.014036192558705807, + 0.018945464864373207, + 0.016283424571156502, + 0.020259084179997444, + 0.012170432135462761, + 0.01162896491587162, + 0.015467238612473011, + 0.03200630843639374, + 0.0037226758431643248, + 0.002177531598135829, + 0.004076485522091389, + 0.003783927299082279, + 0.0019579450599849224, + 0.0024532023817300797, + 0.004062404856085777, + 0.0028542526997625828, + 0.0013807232026010752, + 0.0008821208029985428, + 0.001505075255408883, + 0.001101541449315846, + 0.0008922939305193722, + 0.0013159498339518905, + 0.0010755002731457353, + 0.001301031792536378, + 0.002034364966675639, + 0.0022513337898999453, + 0.0017685415223240852, + 0.0017226130003109574, + 0.0036245761439204216, + 0.0017208565259352326, + 0.002556364517658949, + 0.0020326138474047184, + 0.0011300516780465841, + 0.0008552765939384699, + 0.0015475020045414567, + 0.0009397590765729547, + 0.0011730471160262823, + 0.001736996229737997, + 0.0013611660106107593, + 0.001663020346313715, + 0.00767942238599062, + 0.002278005238622427, + 0.012144128791987896, + 0.009443195536732674, + 0.0018429089104756713, + 0.003770990762859583, + 0.007658190559595823, + 0.03335998207330704, + 0.02501286379992962, + 0.01289212703704834, + 0.2696862816810608, + 0.2518230676651001, + 0.013638314791023731, + 0.011574937961995602, + 0.020897626876831055, + 0.04041197523474693, + 0.002579934662207961, + 0.0023864677641540766, + 0.005016888491809368, + 0.007956280373036861, + 0.0027158576995134354, + 0.0024393564090132713, + 0.0033511200454086065, + 0.0027548742946237326, + 0.0008480155956931412, + 0.0010942191584035754, + 0.0019339080899953842, + 0.0038293818943202496, + 0.000905503227841109, + 0.0007737150881439447, + 0.0013610729947686195, + 0.0011611758964136243, + 0.001901120413094759, + 0.0015027689514681697, + 0.003109478857368231, + 0.005711079109460115, + 0.0016116751357913017, + 0.0021185220684856176, + 0.009767395444214344, + 0.005598332267254591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0024878999684005976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot, captured at a low angle, pans backward and slowly rises as a train advances, ultimately capturing an epic drone-like tracking perspective. In the frame, a dark, weathered steam locomotive, towing several old, rusted iron carriages, jolted violently across a vast, barren plain. Two rusted tracks, with rotting sleepers, stretch across the cracked, earthy earth until they disappear into the horizon. The locomotive's towering smokestack spews erratically, emitting thick, gray-black steam and soot, mixed with dust, leaving a long, dissipating trail behind it. The train heaves and pitches violently on the uneven track, its joints repeatedly stretched and squeezed. Each jolt threatens to derail, a gesture of its tenacious struggle to move forward. The low-hanging sun shines through the gray-yellow clouds in the sky, casting dim and weak light, casting long shadows on the dry weeds and scattered bushes, dyeing the entire scene with a monotonous and depressing warm tone, creating a desolate, lonely and apocalyptic atmosphere full of primitive mechanical power.", + "seed": 20585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_15/index_4307_seed_20585_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": -0.015148130990564824, + "vlm_logprob": -0.015148130990564824, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.015148130990564824, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.015188867226243019 + ], + "opposite_logprob_tokens": [ + -4.515188694000244 + ], + "logprob_reward": -0.015188867226243019, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015188867226243019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1919415444135666, + 0.05556650087237358, + 0.10215234756469727, + 0.03737546131014824, + 0.07313846796751022, + 0.14465898275375366, + 0.054558705538511276, + 0.1052914634346962, + 0.10088850557804108, + 0.13442808389663696 + ], + "credit_map": [ + 0.029172582551836967, + 0.018251406028866768, + 0.026914315298199654, + 0.025131626054644585, + 0.019590727984905243, + 0.021077333018183708, + 0.02940143272280693, + 0.022402111440896988, + 0.006448216270655394, + 0.008577723056077957, + 0.007263044826686382, + 0.0075821345672011375, + 0.004424339160323143, + 0.00543579226359725, + 0.0065670921467244625, + 0.009268155321478844, + 0.010125387459993362, + 0.008237681351602077, + 0.024044577032327652, + 0.01234945934265852, + 0.006510539911687374, + 0.00949739571660757, + 0.01399431936442852, + 0.01739298552274704, + 0.0048050167970359325, + 0.005868877749890089, + 0.00475167203694582, + 0.006023155525326729, + 0.0033474008087068796, + 0.0040527088567614555, + 0.00417445320636034, + 0.00435217609629035, + 0.009222093969583511, + 0.011312607675790787, + 0.007334588561207056, + 0.00939811859279871, + 0.008091824129223824, + 0.010878022760152817, + 0.008207937702536583, + 0.008693268522620201, + 0.015469356440007687, + 0.022867869585752487, + 0.02107851579785347, + 0.022907722741365433, + 0.021691245958209038, + 0.010641971603035927, + 0.016537711024284363, + 0.013464583083987236, + 0.006843863520771265, + 0.00802969466894865, + 0.005988086108118296, + 0.008867966942489147, + 0.005900675896555185, + 0.00691582914441824, + 0.006211692932993174, + 0.005800892133265734, + 0.00902015995234251, + 0.01250061672180891, + 0.013044293969869614, + 0.014655583538115025, + 0.012981447391211987, + 0.01025742944329977, + 0.016916653141379356, + 0.015915285795927048, + 0.00902241375297308, + 0.013415083289146423, + 0.012051017954945564, + 0.011511124670505524, + 0.01844368502497673, + 0.0164424329996109, + 0.01127681415528059, + 0.00872593279927969, + 0.012022258713841438, + 0.012241515330970287, + 0.01854797825217247, + 0.01611492782831192, + 0.01736460253596306, + 0.025237636640667915, + 0.01275387592613697, + 0.020145291462540627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.015188867226243019 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.029808850958943367 + ], + "opposite_logprob_tokens": [ + -3.529808759689331 + ], + "logprob_reward": -0.029808850958943367, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808850958943367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17556451261043549, + 0.07943739742040634, + 0.04804041609168053, + 0.03017144836485386, + 0.04996369406580925, + 0.15433119237422943, + 0.08359677344560623, + 0.11944223195314407, + 0.09623339772224426, + 0.1632189005613327 + ], + "credit_map": [ + 0.01889294944703579, + 0.013649769127368927, + 0.03266604244709015, + 0.027978086844086647, + 0.011254343204200268, + 0.015139403752982616, + 0.03200393170118332, + 0.023979991674423218, + 0.00786972139030695, + 0.00882046390324831, + 0.017082836478948593, + 0.009803485125303268, + 0.004033680073916912, + 0.0065496391616761684, + 0.010836531408131123, + 0.014441043138504028, + 0.004877736326307058, + 0.00469959806650877, + 0.00805892888456583, + 0.006242390722036362, + 0.004075764212757349, + 0.0047688609920442104, + 0.006754077505320311, + 0.008563058450818062, + 0.004202579148113728, + 0.0028537469916045666, + 0.0034951819106936455, + 0.0036536918487399817, + 0.003045520046725869, + 0.0035941435489803553, + 0.003171768970787525, + 0.006154813803732395, + 0.006979221478104591, + 0.007684489246457815, + 0.005265781190246344, + 0.006434767972677946, + 0.00528771011158824, + 0.007532356306910515, + 0.004468670580536127, + 0.006310697644948959, + 0.015051621943712234, + 0.02443806454539299, + 0.024003617465496063, + 0.029231389984488487, + 0.021619999781250954, + 0.01103570219129324, + 0.014779363758862019, + 0.014171426184475422, + 0.009129030629992485, + 0.017758511006832123, + 0.013217662461102009, + 0.015130138956010342, + 0.008009713143110275, + 0.006956818047910929, + 0.006271799560636282, + 0.007123100105673075, + 0.012147657573223114, + 0.018912823870778084, + 0.012958846986293793, + 0.014184040948748589, + 0.01819125935435295, + 0.011350791901350021, + 0.01586989127099514, + 0.015826918184757233, + 0.008929572999477386, + 0.008799676783382893, + 0.016003044322133064, + 0.010824945755302906, + 0.018080195412039757, + 0.016408348456025124, + 0.010231834836304188, + 0.006955781020224094, + 0.012455631047487259, + 0.014747321605682373, + 0.029164480045437813, + 0.02226857654750347, + 0.01706601306796074, + 0.027260463684797287, + 0.015681197866797447, + 0.024575218558311462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.029808850958943367 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.0028211576864123344 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.0028211576864123344, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028211576864123344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19920866191387177, + 0.14800918102264404, + 0.11586181074380875, + 0.07203535735607147, + 0.09272515028715134, + 0.08049958944320679, + 0.06517701596021652, + 0.04933701828122139, + 0.11700932681560516, + 0.06013688072562218 + ], + "credit_map": [ + 0.034454114735126495, + 0.0293126180768013, + 0.0270405150949955, + 0.018904902040958405, + 0.0226286631077528, + 0.024433933198451996, + 0.015915965661406517, + 0.026517949998378754, + 0.01937166228890419, + 0.03252623602747917, + 0.014837902039289474, + 0.01741662248969078, + 0.014403318054974079, + 0.015626413747668266, + 0.016629179939627647, + 0.017197836190462112, + 0.01245648879557848, + 0.009365522302687168, + 0.026129132136702538, + 0.012729857116937637, + 0.010489663109183311, + 0.010787155479192734, + 0.014243245124816895, + 0.019660742953419685, + 0.015092783607542515, + 0.010230091400444508, + 0.00837971642613411, + 0.007819287478923798, + 0.005212439689785242, + 0.0074823638424277306, + 0.006683659739792347, + 0.011135014705359936, + 0.013176481239497662, + 0.008460436016321182, + 0.009466923773288727, + 0.011048237793147564, + 0.010817517526447773, + 0.01603417657315731, + 0.009337267838418484, + 0.014384113252162933, + 0.007520646322518587, + 0.009865774773061275, + 0.014825092628598213, + 0.014906604774296284, + 0.00986889936029911, + 0.006397082936018705, + 0.008286738768219948, + 0.008828744292259216, + 0.008295170962810516, + 0.016900911927223206, + 0.006900035310536623, + 0.006623408291488886, + 0.007539595011621714, + 0.005771793890744448, + 0.00719169806689024, + 0.005954406224191189, + 0.005736213177442551, + 0.00560341402888298, + 0.0048051681369543076, + 0.005214835051447153, + 0.0062439702451229095, + 0.010503722354769707, + 0.0059679411351680756, + 0.0052617574110627174, + 0.010372303426265717, + 0.014109592884778976, + 0.014181575737893581, + 0.014850038103759289, + 0.016804859042167664, + 0.018961317837238312, + 0.01695920154452324, + 0.010770441964268684, + 0.004250956699252129, + 0.007082647178322077, + 0.009854014962911606, + 0.00995427742600441, + 0.005318006034940481, + 0.006928842980414629, + 0.009362847544252872, + 0.007385290693491697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0028211576864123344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.02328469604253769 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.02328469604253769, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328469604253769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09231410175561905, + 0.07329194247722626, + 0.099949412047863, + 0.041460562497377396, + 0.050698600709438324, + 0.18205483257770538, + 0.17565900087356567, + 0.1788375973701477, + 0.06134508550167084, + 0.04438892379403114 + ], + "credit_map": [ + 0.009616547264158726, + 0.011117497459053993, + 0.01600606180727482, + 0.012733378447592258, + 0.009562970139086246, + 0.008063273504376411, + 0.013950048014521599, + 0.011264321394264698, + 0.006650327239185572, + 0.009103670716285706, + 0.012751431204378605, + 0.007494593039155006, + 0.005803164094686508, + 0.006997940596193075, + 0.012422735802829266, + 0.012068080715835094, + 0.009242083877325058, + 0.009731432422995567, + 0.02228229120373726, + 0.01187701802700758, + 0.0067168911918997765, + 0.00933231133967638, + 0.011984365060925484, + 0.01878301613032818, + 0.004647459834814072, + 0.006166039500385523, + 0.004754668101668358, + 0.00579436169937253, + 0.0033547081984579563, + 0.003932345658540726, + 0.004577445797622204, + 0.008233534172177315, + 0.006590938661247492, + 0.009870165959000587, + 0.006021973676979542, + 0.006303905975073576, + 0.005231030751019716, + 0.005501916166394949, + 0.005272333975881338, + 0.005906335543841124, + 0.01405305601656437, + 0.02572617307305336, + 0.04317689314484596, + 0.041374314576387405, + 0.015200566500425339, + 0.009243923239409924, + 0.01629575341939926, + 0.01698414422571659, + 0.01880832016468048, + 0.02883416973054409, + 0.025375446304678917, + 0.020834797993302345, + 0.0355936624109745, + 0.016745949164032936, + 0.016293134540319443, + 0.013173520565032959, + 0.01926233060657978, + 0.01764831505715847, + 0.01624785177409649, + 0.019353248178958893, + 0.03339492902159691, + 0.027531158179044724, + 0.02437727153301239, + 0.021022481843829155, + 0.005188897252082825, + 0.005916254594922066, + 0.00833938643336296, + 0.007743386086076498, + 0.010252032428979874, + 0.00846860371530056, + 0.010000408627092838, + 0.005436118226498365, + 0.003885525744408369, + 0.0048931366764009, + 0.007371037267148495, + 0.0055883196182549, + 0.004591840784996748, + 0.006782026961445808, + 0.004902387969195843, + 0.00637464877218008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.02328469604253769 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.004637083038687706 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637083038687706, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637083038687706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1344309002161026, + 0.09321930259466171, + 0.03224100172519684, + 0.03191899508237839, + 0.07150240242481232, + 0.08535721898078918, + 0.31964191794395447, + 0.14619089663028717, + 0.04338203743100166, + 0.04211529344320297 + ], + "credit_map": [ + 0.01583506353199482, + 0.013390117324888706, + 0.021294455975294113, + 0.02178623527288437, + 0.00994074996560812, + 0.011902201920747757, + 0.023694775998592377, + 0.01658730022609234, + 0.008044776506721973, + 0.014965226873755455, + 0.015687476843595505, + 0.011472324840724468, + 0.006838059984147549, + 0.00798090547323227, + 0.01549770962446928, + 0.012732826173305511, + 0.0032202487345784903, + 0.003966795280575752, + 0.003927214536815882, + 0.004759892821311951, + 0.0037480792962014675, + 0.0036885919980704784, + 0.004841913934797049, + 0.004088267683982849, + 0.0037576614413410425, + 0.003723490284755826, + 0.004869834054261446, + 0.004550442565232515, + 0.0021112544927746058, + 0.0031997112091630697, + 0.004131126683205366, + 0.005575473420321941, + 0.011247972026467323, + 0.006390165537595749, + 0.0077459849417209625, + 0.010968120768666267, + 0.009111186489462852, + 0.011320524848997593, + 0.0068852659314870834, + 0.00783318467438221, + 0.007043265271931887, + 0.010368973016738892, + 0.015176959335803986, + 0.019136125221848488, + 0.007275014650076628, + 0.0061309379525482655, + 0.008924175053834915, + 0.011301768012344837, + 0.0409722663462162, + 0.08830273151397705, + 0.03508449345827103, + 0.038518380373716354, + 0.03488371521234512, + 0.02701479382812977, + 0.03069925121963024, + 0.02416629157960415, + 0.016888132318854332, + 0.01705101504921913, + 0.025046424940228462, + 0.022334666922688484, + 0.017357949167490005, + 0.020937299355864525, + 0.01228801067918539, + 0.014287394471466541, + 0.003935185261070728, + 0.003916047979146242, + 0.004507783334702253, + 0.004748398903757334, + 0.008891066536307335, + 0.0061484891921281815, + 0.007202307227998972, + 0.004032759461551905, + 0.0033584577031433582, + 0.0034976950846612453, + 0.006525197997689247, + 0.006093122065067291, + 0.004229324404150248, + 0.007656069006770849, + 0.005193987861275673, + 0.005561437923461199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.004637083038687706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot, captured at a low angle, pans backward and slowly rises as a train advances, ultimately capturing an epic drone-like tracking perspective. In the frame, a dark, weathered steam locomotive, towing several old, rusted iron carriages, jolted violently across a vast, barren plain. Two rusted tracks, with rotting sleepers, stretch across the cracked, earthy earth until they disappear into the horizon. The locomotive's towering smokestack spews erratically, emitting thick, gray-black steam and soot, mixed with dust, leaving a long, dissipating trail behind it. The train heaves and pitches violently on the uneven track, its joints repeatedly stretched and squeezed. Each jolt threatens to derail, a gesture of its tenacious struggle to move forward. The low-hanging sun shines through the gray-yellow clouds in the sky, casting dim and weak light, casting long shadows on the dry weeds and scattered bushes, dyeing the entire scene with a monotonous and depressing warm tone, creating a desolate, lonely and apocalyptic atmosphere full of primitive mechanical power.", + "seed": 20586, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_15/index_4307_seed_20586_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": -0.015130834840238094, + "vlm_logprob": -0.015130834840238094, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.015130834840238094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.017811819911003113 + ], + "opposite_logprob_tokens": [ + -4.3928117752075195 + ], + "logprob_reward": -0.017811819911003113, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017811819911003113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23795437812805176, + 0.09893698990345001, + 0.14365971088409424, + 0.13893388211727142, + 0.10264576971530914, + 0.04822315275669098, + 0.04597147926688194, + 0.03162563219666481, + 0.12230004370212555, + 0.02974897436797619 + ], + "credit_map": [ + 0.025801178067922592, + 0.04920099675655365, + 0.022161681205034256, + 0.02580101601779461, + 0.0353284627199173, + 0.030097154900431633, + 0.021744204685091972, + 0.02781970612704754, + 0.011369760148227215, + 0.013250494375824928, + 0.016633829101920128, + 0.011022379621863365, + 0.009572576731443405, + 0.010724595747888088, + 0.011031161062419415, + 0.015332196839153767, + 0.01668313704431057, + 0.018228601664304733, + 0.015340020880103111, + 0.029391421005129814, + 0.014070735312998295, + 0.013751287013292313, + 0.019686667248606682, + 0.016507843509316444, + 0.015499516390264034, + 0.013162638060748577, + 0.02051825448870659, + 0.013498803600668907, + 0.038937054574489594, + 0.013294980861246586, + 0.012812747620046139, + 0.011209901422262192, + 0.011147958226501942, + 0.015708666294813156, + 0.012248840183019638, + 0.011712643317878246, + 0.01794763095676899, + 0.01187992375344038, + 0.012211226858198643, + 0.009788893163204193, + 0.003976161126047373, + 0.006858809385448694, + 0.005630189087241888, + 0.004917045123875141, + 0.009955492801964283, + 0.00507896114140749, + 0.006272728089243174, + 0.005533767864108086, + 0.005679067689925432, + 0.006624033208936453, + 0.007880558259785175, + 0.003641432849690318, + 0.0060995835810899734, + 0.005811075214296579, + 0.005501683801412582, + 0.004734047222882509, + 0.0035635316744446754, + 0.004863786045461893, + 0.0036718370392918587, + 0.0033825726713985205, + 0.003963829018175602, + 0.005693119019269943, + 0.003389362245798111, + 0.003097597975283861, + 0.013736095279455185, + 0.012411346659064293, + 0.01164817251265049, + 0.009376736357808113, + 0.017025211825966835, + 0.021628333255648613, + 0.01971323974430561, + 0.016760922968387604, + 0.0037230467423796654, + 0.004153164569288492, + 0.0026359667535871267, + 0.002146578161045909, + 0.003386878641322255, + 0.0034099924378097057, + 0.005085519514977932, + 0.005207829177379608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.017811819911003113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.016076110303401947 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016076110303401947, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076110303401947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23452956974506378, + 0.10640288889408112, + 0.1135319247841835, + 0.21361562609672546, + 0.1175873652100563, + 0.05045272037386894, + 0.04169908165931702, + 0.04035697132349014, + 0.05831329524517059, + 0.023510552942752838 + ], + "credit_map": [ + 0.022196169942617416, + 0.057070862501859665, + 0.0180456954985857, + 0.02136559970676899, + 0.036958061158657074, + 0.03113410808146, + 0.022689832374453545, + 0.025069231167435646, + 0.013071720488369465, + 0.014143701642751694, + 0.016802966594696045, + 0.010316469706594944, + 0.012526096776127815, + 0.011686250567436218, + 0.011032775044441223, + 0.016822898760437965, + 0.013898882083594799, + 0.014792114496231079, + 0.012495497241616249, + 0.024661537259817123, + 0.010768432170152664, + 0.009515895508229733, + 0.013612240552902222, + 0.013787321746349335, + 0.024525653570890427, + 0.026125244796276093, + 0.01794498600065708, + 0.015251867473125458, + 0.06552420556545258, + 0.025421831756830215, + 0.021493472158908844, + 0.017328374087810516, + 0.01095473300665617, + 0.0197871346026659, + 0.015048916451632977, + 0.01647358201444149, + 0.022619344294071198, + 0.013149470090866089, + 0.010122662410140038, + 0.00943153165280819, + 0.004708569031208754, + 0.008032312616705894, + 0.004782856907695532, + 0.004221066832542419, + 0.013286031782627106, + 0.007343650329858065, + 0.004150400869548321, + 0.003927834797650576, + 0.005052158143371344, + 0.0084474952891469, + 0.006163629703223705, + 0.003716825507581234, + 0.004679868463426828, + 0.006116779055446386, + 0.0031760274432599545, + 0.004346298053860664, + 0.003895105794072151, + 0.00426270579919219, + 0.006606692913919687, + 0.005764335859566927, + 0.004535735119134188, + 0.004712430294603109, + 0.005150620825588703, + 0.005429345648735762, + 0.0065778326243162155, + 0.011354057118296623, + 0.006189711391925812, + 0.007315626367926598, + 0.006800761912018061, + 0.006565274205058813, + 0.007560147903859615, + 0.005949883256107569, + 0.002811112441122532, + 0.002833848586305976, + 0.002302595181390643, + 0.0024934124667197466, + 0.0020113191567361355, + 0.0028930669650435448, + 0.003673689905554056, + 0.004491508938372135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.016076110303401947 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12821988761425018, + 0.04446909576654434, + 0.09181241691112518, + 0.29906946420669556, + 0.07981664687395096, + 0.14628079533576965, + 0.07274603098630905, + 0.07259466499090195, + 0.041854023933410645, + 0.023136910051107407 + ], + "credit_map": [ + 0.012737125158309937, + 0.03160158172249794, + 0.010933059267699718, + 0.01331829559057951, + 0.01888238452374935, + 0.016550587490200996, + 0.011639364063739777, + 0.0125574991106987, + 0.00415608799085021, + 0.0072762263007462025, + 0.004561638925224543, + 0.004729894455522299, + 0.005966047756373882, + 0.005782861262559891, + 0.004982329439371824, + 0.007014008238911629, + 0.009046973660588264, + 0.02164987102150917, + 0.005965440068393946, + 0.013340942561626434, + 0.01519801001995802, + 0.011166463606059551, + 0.008763757534325123, + 0.006680958438664675, + 0.033262815326452255, + 0.03743279352784157, + 0.026074446737766266, + 0.021672025322914124, + 0.09711306542158127, + 0.03254638612270355, + 0.026636909693479538, + 0.02433103509247303, + 0.008071478456258774, + 0.010321743786334991, + 0.008143842220306396, + 0.01816612109541893, + 0.008884193375706673, + 0.006301966495811939, + 0.012697432190179825, + 0.007229866459965706, + 0.010835445486009121, + 0.023944519460201263, + 0.010758474469184875, + 0.009117376990616322, + 0.05478281155228615, + 0.013530580326914787, + 0.013814653269946575, + 0.00949694775044918, + 0.007677013520151377, + 0.011554700322449207, + 0.008870413526892662, + 0.006153654307126999, + 0.008768357336521149, + 0.014977221377193928, + 0.007555045187473297, + 0.007189628202468157, + 0.006268370896577835, + 0.007258334197103977, + 0.006637233309447765, + 0.007118349429219961, + 0.006163752172142267, + 0.012141173705458641, + 0.016928281635046005, + 0.01007916871458292, + 0.004137715790420771, + 0.004481226671487093, + 0.004544479306787252, + 0.004336822312325239, + 0.004398687742650509, + 0.005785322282463312, + 0.00804099626839161, + 0.006128772161900997, + 0.0026317816227674484, + 0.001908050267957151, + 0.0020069279707968235, + 0.0015196821186691523, + 0.0021942281164228916, + 0.0027577218133956194, + 0.004507799632847309, + 0.005610717460513115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.005255098920315504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.0336914137005806 + ], + "opposite_logprob_tokens": [ + -3.40869140625 + ], + "logprob_reward": -0.0336914137005806, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336914137005806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15013468265533447, + 0.08183422684669495, + 0.045352865010499954, + 0.20432257652282715, + 0.08351857960224152, + 0.13948538899421692, + 0.07961210608482361, + 0.10261739790439606, + 0.07802386581897736, + 0.03509829193353653 + ], + "credit_map": [ + 0.016994884237647057, + 0.03567451238632202, + 0.011966373771429062, + 0.014529146254062653, + 0.023524774238467216, + 0.019639182835817337, + 0.01322111114859581, + 0.014584698714315891, + 0.014077159576117992, + 0.011511974036693573, + 0.011610143817961216, + 0.008296863175928593, + 0.007579905912280083, + 0.00996629148721695, + 0.007255330681800842, + 0.011536569334566593, + 0.005594268441200256, + 0.007318345829844475, + 0.004215151537209749, + 0.005950035061687231, + 0.008018376305699348, + 0.005144539289176464, + 0.0048269485123455524, + 0.004285203292965889, + 0.022377988323569298, + 0.025909869000315666, + 0.021440356969833374, + 0.01714823581278324, + 0.06256788969039917, + 0.01801072433590889, + 0.020448753610253334, + 0.016418786719441414, + 0.009219657629728317, + 0.01342227403074503, + 0.010660037398338318, + 0.008971331641077995, + 0.013168949633836746, + 0.009958037175238132, + 0.01075445581227541, + 0.007363845128566027, + 0.012670652940869331, + 0.018560033291578293, + 0.013957369141280651, + 0.010583621449768543, + 0.024775506928563118, + 0.03164595365524292, + 0.012922050431370735, + 0.014370210468769073, + 0.008150370791554451, + 0.01363674271851778, + 0.00874308217316866, + 0.005164485424757004, + 0.011399378068745136, + 0.0170089490711689, + 0.008026378229260445, + 0.007482725661247969, + 0.008946246467530727, + 0.014271614141762257, + 0.010388371534645557, + 0.013677116483449936, + 0.009544764645397663, + 0.012749963440001011, + 0.017908452078700066, + 0.015130875632166862, + 0.008689775131642818, + 0.011328082531690598, + 0.006994002964347601, + 0.0067444415763020515, + 0.009028441272675991, + 0.012995832599699497, + 0.01196443010121584, + 0.010278868488967419, + 0.004869070369750261, + 0.004198012407869101, + 0.003358658403158188, + 0.0028782354202121496, + 0.0036348318681120872, + 0.005585179198533297, + 0.006461568176746368, + 0.004112740512937307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0336914137005806 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11944549530744553, + 0.08430945128202438, + 0.18828877806663513, + 0.08190987259149551, + 0.15246129035949707, + 0.09940636903047562, + 0.07560968399047852, + 0.0619131401181221, + 0.08114782720804214, + 0.0555080883204937 + ], + "credit_map": [ + 0.011276571080088615, + 0.021092260256409645, + 0.01007151324301958, + 0.016354432329535484, + 0.016865873709321022, + 0.014230205677449703, + 0.011814123950898647, + 0.0177405197173357, + 0.010471822693943977, + 0.010838967747986317, + 0.011749284341931343, + 0.008395415730774403, + 0.010611520148813725, + 0.009590139612555504, + 0.009816458448767662, + 0.012835849076509476, + 0.02526833675801754, + 0.04512180760502815, + 0.014208389446139336, + 0.012680496089160442, + 0.031869515776634216, + 0.025716999545693398, + 0.017026226967573166, + 0.01639699935913086, + 0.010157991200685501, + 0.012789213098585606, + 0.008000526577234268, + 0.007817969657480717, + 0.014126843772828579, + 0.011725719086825848, + 0.010189812630414963, + 0.0071017928421497345, + 0.015103806741535664, + 0.024258676916360855, + 0.021297862753272057, + 0.021965494379401207, + 0.02579224854707718, + 0.016639793291687965, + 0.013693257234990597, + 0.013710148632526398, + 0.009329967200756073, + 0.018984122201800346, + 0.012485126033425331, + 0.011475332081317902, + 0.013492636382579803, + 0.011951787397265434, + 0.009096152149140835, + 0.01259124930948019, + 0.008299910463392735, + 0.010909470729529858, + 0.009984483942389488, + 0.007789934519678354, + 0.00892968475818634, + 0.017325930297374725, + 0.007029201835393906, + 0.005341066047549248, + 0.005035300273448229, + 0.006464214995503426, + 0.007544000633060932, + 0.007460079621523619, + 0.006471527274698019, + 0.00902239978313446, + 0.010672610253095627, + 0.009243009611964226, + 0.008430692367255688, + 0.014529402367770672, + 0.009277235716581345, + 0.011530128307640553, + 0.007680011913180351, + 0.010047207586467266, + 0.009078053757548332, + 0.01057509332895279, + 0.007299886085093021, + 0.00965121854096651, + 0.004925600253045559, + 0.005775188561528921, + 0.006389658898115158, + 0.008040022104978561, + 0.007231943774968386, + 0.006194571498781443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.002819731365889311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide, long shot, captured at a low angle, pans backward and slowly rises as a train advances, ultimately capturing an epic drone-like tracking perspective. In the frame, a dark, weathered steam locomotive, towing several old, rusted iron carriages, jolted violently across a vast, barren plain. Two rusted tracks, with rotting sleepers, stretch across the cracked, earthy earth until they disappear into the horizon. The locomotive's towering smokestack spews erratically, emitting thick, gray-black steam and soot, mixed with dust, leaving a long, dissipating trail behind it. The train heaves and pitches violently on the uneven track, its joints repeatedly stretched and squeezed. Each jolt threatens to derail, a gesture of its tenacious struggle to move forward. The low-hanging sun shines through the gray-yellow clouds in the sky, casting dim and weak light, casting long shadows on the dry weeds and scattered bushes, dyeing the entire scene with a monotonous and depressing warm tone, creating a desolate, lonely and apocalyptic atmosphere full of primitive mechanical power.", + "seed": 20587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_15/index_4307_seed_20587_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": -0.02366819861344993, + "vlm_logprob": -0.02366819861344993, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02366819861344993, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.042186420410871506 + ], + "opposite_logprob_tokens": [ + -3.2921864986419678 + ], + "logprob_reward": -0.042186420410871506, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.042186420410871506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33114495873451233, + 0.06906136125326157, + 0.09938301146030426, + 0.040181998163461685, + 0.06049074977636337, + 0.08410795778036118, + 0.05535242706537247, + 0.07434765994548798, + 0.056488052010536194, + 0.12944182753562927 + ], + "credit_map": [ + 0.09207410365343094, + 0.029749104753136635, + 0.010792747139930725, + 0.015611768700182438, + 0.0385182648897171, + 0.09935121238231659, + 0.016726521775126457, + 0.02832123450934887, + 0.017211223021149635, + 0.008632722310721874, + 0.0031857818830758333, + 0.005239794030785561, + 0.007753672543913126, + 0.015600113198161125, + 0.004947389476001263, + 0.006490666884928942, + 0.03096204437315464, + 0.009070640429854393, + 0.004294403828680515, + 0.0058901249431073666, + 0.007997897453606129, + 0.0319063737988472, + 0.004463841672986746, + 0.004797681234776974, + 0.006776297464966774, + 0.004049406386911869, + 0.003140719374641776, + 0.004588826093822718, + 0.005965495947748423, + 0.00843468215316534, + 0.0035070888698101044, + 0.00371947861276567, + 0.006657900754362345, + 0.011023436672985554, + 0.004560600500553846, + 0.00970451906323433, + 0.011081642471253872, + 0.00619293749332428, + 0.0066162412986159325, + 0.00465346546843648, + 0.020622195675969124, + 0.008564424701035023, + 0.004925713408738375, + 0.006994013208895922, + 0.00642103748396039, + 0.023417510092258453, + 0.006465589627623558, + 0.006697477772831917, + 0.0067896670661866665, + 0.005925155244767666, + 0.003835804993286729, + 0.007938439026474953, + 0.007445891387760639, + 0.008141747675836086, + 0.009816155768930912, + 0.0054595619440078735, + 0.009435662999749184, + 0.008266585879027843, + 0.006044994108378887, + 0.013017942197620869, + 0.011261148378252983, + 0.008258713409304619, + 0.01210823655128479, + 0.00595436803996563, + 0.004880256485193968, + 0.004437707830220461, + 0.005815951619297266, + 0.007162956986576319, + 0.017000455409288406, + 0.00521421292796731, + 0.007517256308346987, + 0.004459254443645477, + 0.012639230117201805, + 0.014753182418644428, + 0.015604117885231972, + 0.01693606562912464, + 0.01433129608631134, + 0.016866039484739304, + 0.013792159967124462, + 0.024519730359315872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.042186420410871506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.029807347804307938 + ], + "opposite_logprob_tokens": [ + -3.5298073291778564 + ], + "logprob_reward": -0.029807347804307938, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807347804307938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35937249660491943, + 0.19431531429290771, + 0.12771138548851013, + 0.022830994799733162, + 0.06716499477624893, + 0.03733767941594124, + 0.06281875818967819, + 0.02419763244688511, + 0.0564422570168972, + 0.04780846834182739 + ], + "credit_map": [ + 0.1146988794207573, + 0.017370296642184258, + 0.011334928683936596, + 0.009151865728199482, + 0.03422481566667557, + 0.13681843876838684, + 0.013059702701866627, + 0.02271357551217079, + 0.05724053829908371, + 0.007446896750479937, + 0.0063762725330889225, + 0.005017003510147333, + 0.022135423496365547, + 0.07421023398637772, + 0.008336511440575123, + 0.013552450574934483, + 0.04374253377318382, + 0.006318596191704273, + 0.005116446875035763, + 0.005260145757347345, + 0.007271707523614168, + 0.047534629702568054, + 0.005508346017450094, + 0.006958981975913048, + 0.0037415807601064444, + 0.0024476335383951664, + 0.0021206841338425875, + 0.002800802933052182, + 0.0030141875613480806, + 0.004016312770545483, + 0.0019361324375495315, + 0.0027536607813090086, + 0.009096290916204453, + 0.012767929583787918, + 0.005085616372525692, + 0.007864603772759438, + 0.012836464680731297, + 0.009012577123939991, + 0.005328613333404064, + 0.005172898527234793, + 0.008421618491411209, + 0.004092911258339882, + 0.002237123902887106, + 0.003388546174392104, + 0.003222260857000947, + 0.009682591073215008, + 0.0029272923711687326, + 0.003365333890542388, + 0.018971135839819908, + 0.0031302962452173233, + 0.003159670624881983, + 0.005750236101448536, + 0.003330915467813611, + 0.02089458890259266, + 0.0044255671091377735, + 0.0031563458032906055, + 0.002764117205515504, + 0.002350238151848316, + 0.0026236025150865316, + 0.004155179485678673, + 0.003563490929082036, + 0.003179088467732072, + 0.003327780868858099, + 0.0022341355215758085, + 0.0050670308992266655, + 0.005000874865800142, + 0.004383048042654991, + 0.0077543240040540695, + 0.017024196684360504, + 0.005564059130847454, + 0.007391241379082203, + 0.004257481545209885, + 0.00466917222365737, + 0.004644425120204687, + 0.004146276973187923, + 0.006864613387733698, + 0.012729346752166748, + 0.005842786747962236, + 0.005602716002613306, + 0.0033091299701482058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029807347804307938 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.004640643019229174 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640643019229174, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640643019229174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21861083805561066, + 0.057779137045145035, + 0.1575022041797638, + 0.01796562224626541, + 0.2157631367444992, + 0.08473903685808182, + 0.13866694271564484, + 0.039124492555856705, + 0.04682439565658569, + 0.02302420139312744 + ], + "credit_map": [ + 0.05966849997639656, + 0.013233590871095657, + 0.009554170072078705, + 0.015932613983750343, + 0.023147987201809883, + 0.07322349399328232, + 0.008418012410402298, + 0.015432474203407764, + 0.015727175399661064, + 0.005311930552124977, + 0.0022262202110141516, + 0.005373202729970217, + 0.005142367444932461, + 0.01784515753388405, + 0.0025743378791958094, + 0.003578740172088146, + 0.05572992563247681, + 0.010973615571856499, + 0.005345470737665892, + 0.006845957599580288, + 0.008949203416705132, + 0.05928520858287811, + 0.005222688429057598, + 0.005150139331817627, + 0.002712198533117771, + 0.0018934155814349651, + 0.0015254843747243285, + 0.0019395690178498626, + 0.002711524022743106, + 0.002850595163181424, + 0.0013593640178442001, + 0.0029734699055552483, + 0.05749650299549103, + 0.029737252742052078, + 0.008603996597230434, + 0.01256175059825182, + 0.02865612506866455, + 0.06004875898361206, + 0.010503167286515236, + 0.008155575022101402, + 0.02119518630206585, + 0.00868684146553278, + 0.006203461904078722, + 0.005164104979485273, + 0.006079873535782099, + 0.02356579154729843, + 0.008118689991533756, + 0.005725089460611343, + 0.052215274423360825, + 0.004948427900671959, + 0.0034703209530562162, + 0.004632133524864912, + 0.0060750930570065975, + 0.05696793645620346, + 0.006141677498817444, + 0.00421609403565526, + 0.004570627585053444, + 0.004180859308689833, + 0.004324423149228096, + 0.006794641725718975, + 0.0063551850616931915, + 0.004332820884883404, + 0.005465363617986441, + 0.0031005681958049536, + 0.003934764303267002, + 0.004113044124096632, + 0.004832158330827951, + 0.005114669445902109, + 0.014440099708735943, + 0.0049543618224561214, + 0.00574879813939333, + 0.0036865000147372484, + 0.0021296078339219093, + 0.002213162137195468, + 0.0017631041118875146, + 0.003003582824021578, + 0.00543825700879097, + 0.00279714516364038, + 0.003190513700246811, + 0.0024888277985155582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.004640643019229174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.03809202089905739 + ], + "opposite_logprob_tokens": [ + -3.2880921363830566 + ], + "logprob_reward": -0.03809202089905739, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809202089905739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16382215917110443, + 0.057497479021549225, + 0.1247316524386406, + 0.024787744507193565, + 0.23245204985141754, + 0.06074485182762146, + 0.06513640284538269, + 0.07287316024303436, + 0.1177026703953743, + 0.08025182783603668 + ], + "credit_map": [ + 0.049318891018629074, + 0.015010387636721134, + 0.006978178396821022, + 0.009096423164010048, + 0.01465061865746975, + 0.04875975474715233, + 0.00840816181153059, + 0.011599745601415634, + 0.014215530827641487, + 0.0058365194126963615, + 0.003445280948653817, + 0.003653604304417968, + 0.00633099302649498, + 0.01393105648458004, + 0.004727098625153303, + 0.005357387475669384, + 0.0447547510266304, + 0.004360065329819918, + 0.00435442803427577, + 0.0032727052457630634, + 0.006987248547375202, + 0.05020596832036972, + 0.004897300619632006, + 0.005899177864193916, + 0.004657261073589325, + 0.0027068164199590683, + 0.0016153324395418167, + 0.0029500573873519897, + 0.003257104894146323, + 0.004439478274434805, + 0.0022692084312438965, + 0.0028924839571118355, + 0.02848011441528797, + 0.046591684222221375, + 0.018275130540132523, + 0.027569428086280823, + 0.04737352579832077, + 0.02715863287448883, + 0.019186511635780334, + 0.01781700737774372, + 0.007063279394060373, + 0.009148068726062775, + 0.006359674036502838, + 0.009511664509773254, + 0.007183426525443792, + 0.007294127717614174, + 0.007721117231994867, + 0.006463487632572651, + 0.011224939487874508, + 0.005800528917461634, + 0.005966005381196737, + 0.00808034185320139, + 0.007201396394520998, + 0.011797521263360977, + 0.009459053166210651, + 0.005606614984571934, + 0.008768005296587944, + 0.009167719632387161, + 0.007942572236061096, + 0.011705626733601093, + 0.01220915000885725, + 0.006040581036359072, + 0.010082685388624668, + 0.006956824567168951, + 0.009045676328241825, + 0.008591261692345142, + 0.01266250666230917, + 0.014011966995894909, + 0.03737357631325722, + 0.009797099977731705, + 0.016694702208042145, + 0.009525877423584461, + 0.007025137078016996, + 0.011706196703016758, + 0.007000618614256382, + 0.011116414330899715, + 0.015182692557573318, + 0.008938344195485115, + 0.010903793387115002, + 0.008378621190786362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.03809202089905739 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a dark, weathered steam locomotive towing old, rusted iron carriages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving across a vast, barren plain with cracked earth and rotting sleepers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the locomotive emit thick, gray-black steam and soot from its towering smokestack?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot wide and long, starting at a low angle and panning backward while slowly rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a low-hanging sun through gray-yellow clouds, casting a monotonous warm tone and long 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": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17470420897006989, + 0.03959045559167862, + 0.24295969307422638, + 0.08815848082304001, + 0.10719236731529236, + 0.04145267233252525, + 0.0942382887005806, + 0.12372289597988129, + 0.059526510536670685, + 0.028454480692744255 + ], + "credit_map": [ + 0.021578993648290634, + 0.01747845858335495, + 0.019802995026111603, + 0.013202043250203133, + 0.02809617482125759, + 0.017373403534293175, + 0.020277295261621475, + 0.036894842982292175, + 0.00402915058657527, + 0.005490610376000404, + 0.004365864675492048, + 0.0036716721951961517, + 0.004844196606427431, + 0.004208216909319162, + 0.005487484857439995, + 0.007493257988244295, + 0.08794602006673813, + 0.008755233138799667, + 0.007758208550512791, + 0.007369214203208685, + 0.014685239642858505, + 0.09886886924505234, + 0.0074132406152784824, + 0.010163671337068081, + 0.01145701389759779, + 0.010145590640604496, + 0.010015871375799179, + 0.012701467610895634, + 0.012337432242929935, + 0.013566559180617332, + 0.00905679538846016, + 0.008877748623490334, + 0.033911582082509995, + 0.008605421520769596, + 0.004879536107182503, + 0.004051616415381432, + 0.00981462188065052, + 0.03682628646492958, + 0.0038338073063641787, + 0.0052694943733513355, + 0.00849244650453329, + 0.0040802014991641045, + 0.004385421052575111, + 0.003362855175510049, + 0.0030467172618955374, + 0.009463881142437458, + 0.004214542917907238, + 0.004406608175486326, + 0.02786199375987053, + 0.004515731707215309, + 0.004521917086094618, + 0.007287563290446997, + 0.006737960502505302, + 0.03122301958501339, + 0.006303192116320133, + 0.00578690879046917, + 0.017439592629671097, + 0.019345426931977272, + 0.01045202650129795, + 0.01247345469892025, + 0.013817451894283295, + 0.025831321254372597, + 0.011218199506402016, + 0.013145416975021362, + 0.004763290751725435, + 0.004933692514896393, + 0.006065200548619032, + 0.006704969797283411, + 0.01865990459918976, + 0.005653228610754013, + 0.007954735308885574, + 0.004791487939655781, + 0.0030376084614545107, + 0.0034563520457595587, + 0.003236509393900633, + 0.005435599945485592, + 0.004129162523895502, + 0.003205162938684225, + 0.003273371607065201, + 0.002680713776499033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.0036145609337836504 + } + ], + "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_2/index_tensor([3272, 3272])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_tensor([3272, 3272])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d591941a6c7d4bc4df55381e62d3c49f390bb8e3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_tensor([3272, 3272])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium close-up shot at eye level captures a scene rendered in a vibrant, cheerful, cartoon-like comic book art style, defined by clean lines and bright, saturated colors. A young adult man with warm, olive skin and unkempt black hair is the focal point, dressed in a simple white T-shirt and a sky-blue apron. He leans slightly over a modern stainless steel stovetop, where a cherry-red enameled pot gently simmers its contents, a rich, warm orange. The entire kitchen is bathed in bright, even overhead lighting, making every color pop and eliminating harsh shadows. He carefully lifts a wooden tasting spoon from the pot and steadily brings it to his mouth, delicate wisps of steam curling upward. His expression is a perfect blend of concentration and delight—his brows furrowed slightly, but his eyes sparkle as he carefully savors the liquid, his lips curled in a subtle, satisfied smile, capturing the precise, joyful moment of perfecting his recipe in a warm and wholesome atmosphere.", + "seed": 23637, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_3272_seed_23637_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": -0.015157061722129583, + "vlm_logprob": -0.015157061722129583, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.015157061722129583, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00335119292140007 + ], + "opposite_logprob_tokens": [ + -8.753351211547852 + ], + "logprob_reward": -0.00335119292140007, + "yes_no_margin_reward": 8.75, + "token_credit_reward": -0.00335119292140007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21930333971977234, + 0.08818050473928452, + 0.07338510453701019, + 0.06365210562944412, + 0.03412684053182602, + 0.151365727186203, + 0.15889257192611694, + 0.06719951331615448, + 0.07241565734148026, + 0.07147864997386932 + ], + "credit_map": [ + 0.040808919817209244, + 0.04542812705039978, + 0.026915963739156723, + 0.022687971591949463, + 0.019861185923218727, + 0.01377736497670412, + 0.019228965044021606, + 0.030594825744628906, + 0.013432608917355537, + 0.014228102751076221, + 0.013956919312477112, + 0.012211902067065239, + 0.00716389948502183, + 0.00654543237760663, + 0.006677459459751844, + 0.013964167796075344, + 0.011479569599032402, + 0.009252971038222313, + 0.013740004040300846, + 0.010603610426187515, + 0.00908888690173626, + 0.0038382059428840876, + 0.006176202557981014, + 0.009205646812915802, + 0.01040747668594122, + 0.009903348982334137, + 0.010733090341091156, + 0.009090392850339413, + 0.005457644816488028, + 0.0033971001394093037, + 0.005046224687248468, + 0.009616825729608536, + 0.005307067651301622, + 0.0036352132447063923, + 0.00516151450574398, + 0.004880890250205994, + 0.004194296896457672, + 0.0018213428556919098, + 0.004444204270839691, + 0.004682306200265884, + 0.021684549748897552, + 0.02143927663564682, + 0.03239351138472557, + 0.017747953534126282, + 0.014436054043471813, + 0.010128977708518505, + 0.010718202218413353, + 0.022817207500338554, + 0.02465435490012169, + 0.01395870465785265, + 0.027349287644028664, + 0.02025594748556614, + 0.015272875316441059, + 0.013288984075188637, + 0.016137614846229553, + 0.027974789962172508, + 0.011401080526411533, + 0.013532434590160847, + 0.008802307769656181, + 0.011022308841347694, + 0.006240624934434891, + 0.003476785495877266, + 0.005662574432790279, + 0.00706139812245965, + 0.014187656342983246, + 0.009702938608825207, + 0.014807544648647308, + 0.010929037816822529, + 0.004195480607450008, + 0.003548708511516452, + 0.004972143564373255, + 0.01007214467972517, + 0.011699743568897247, + 0.01112344954162836, + 0.011793093755841255, + 0.008231431245803833, + 0.00898660160601139, + 0.004593694116920233, + 0.006154465954750776, + 0.008896170184016228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.00335119292140007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090751986950636 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.004090751986950636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090751986950636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4118356704711914, + 0.12768825888633728, + 0.03884546086192131, + 0.05022556707262993, + 0.07704312354326248, + 0.08571194857358932, + 0.0492851585149765, + 0.05996754392981529, + 0.06046118959784508, + 0.03893608599901199 + ], + "credit_map": [ + 0.06293931603431702, + 0.071934774518013, + 0.05993939936161041, + 0.047971274703741074, + 0.04097267985343933, + 0.034865736961364746, + 0.03651665523648262, + 0.056695837527513504, + 0.02100777067244053, + 0.019081559032201767, + 0.018109995871782303, + 0.017042044550180435, + 0.011904551647603512, + 0.013421265408396721, + 0.011317823082208633, + 0.01580323837697506, + 0.006539208814501762, + 0.0047001829370856285, + 0.004124074708670378, + 0.005808103363960981, + 0.005639882292598486, + 0.0030515422113239765, + 0.004474412649869919, + 0.004508053418248892, + 0.010076801292598248, + 0.007191127166152, + 0.005681972485035658, + 0.007016139104962349, + 0.0043725380674004555, + 0.005313368048518896, + 0.003925576340407133, + 0.006648044101893902, + 0.011469254270195961, + 0.016256825998425484, + 0.007497256621718407, + 0.008131294511258602, + 0.008377320133149624, + 0.004856592509895563, + 0.00756323616951704, + 0.012891339138150215, + 0.01244064699858427, + 0.00692443223670125, + 0.01386459544301033, + 0.012944311834871769, + 0.009704927913844585, + 0.005551586858928204, + 0.009526140056550503, + 0.014755305834114552, + 0.006812687963247299, + 0.004967838525772095, + 0.007613470312207937, + 0.007308466825634241, + 0.005145591218024492, + 0.006969530135393143, + 0.00528401043266058, + 0.005183563567698002, + 0.009738038294017315, + 0.008484521880745888, + 0.008624442853033543, + 0.008160969242453575, + 0.008364371955394745, + 0.004411625675857067, + 0.00525065278634429, + 0.006932917982339859, + 0.009216857142746449, + 0.005563792772591114, + 0.006428041961044073, + 0.008520626462996006, + 0.009574483148753643, + 0.0043778144754469395, + 0.009074470028281212, + 0.007705105468630791, + 0.005936384201049805, + 0.006748950108885765, + 0.004167559091001749, + 0.0042486912570893764, + 0.004545021336525679, + 0.003303459845483303, + 0.003555895062163472, + 0.006430125329643488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004090751986950636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003610759973526001 + ], + "opposite_logprob_tokens": [ + -5.628610610961914 + ], + "logprob_reward": -0.003610759973526001, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003610759973526001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39023664593696594, + 0.07193630933761597, + 0.06959638744592667, + 0.07340972125530243, + 0.0884900763630867, + 0.1315116584300995, + 0.06500982493162155, + 0.03786172717809677, + 0.038525812327861786, + 0.03342185914516449 + ], + "credit_map": [ + 0.07053269445896149, + 0.1045636385679245, + 0.0426877960562706, + 0.05620892718434334, + 0.024779807776212692, + 0.02147233672440052, + 0.024661464616656303, + 0.0453299805521965, + 0.011891014873981476, + 0.010615121573209763, + 0.009488320909440517, + 0.010569504462182522, + 0.006143732462078333, + 0.005602525547146797, + 0.007975724525749683, + 0.00965036079287529, + 0.011823158711194992, + 0.016275666654109955, + 0.008298518136143684, + 0.00634266622364521, + 0.006262669805437326, + 0.003934428095817566, + 0.004694222006946802, + 0.011965052224695683, + 0.013369154185056686, + 0.019179554656147957, + 0.008636604994535446, + 0.0058125089854002, + 0.008640491403639317, + 0.006159868091344833, + 0.00489899842068553, + 0.0067125423811376095, + 0.01626218855381012, + 0.020741000771522522, + 0.010785216465592384, + 0.008459266275167465, + 0.00571006815880537, + 0.00672321580350399, + 0.005412389058619738, + 0.014396731741726398, + 0.01934599131345749, + 0.013153037056326866, + 0.021215766668319702, + 0.019864438101649284, + 0.014274220913648605, + 0.008417733944952488, + 0.016600092872977257, + 0.018640385940670967, + 0.010590359568595886, + 0.006094914395362139, + 0.007022228557616472, + 0.007589665707200766, + 0.0070418999530375, + 0.0076085957698524, + 0.007144208066165447, + 0.011917955242097378, + 0.005832221824675798, + 0.00885192770510912, + 0.00612702826038003, + 0.005407774355262518, + 0.002535365056246519, + 0.002101100515574217, + 0.0026604041922837496, + 0.00434590270742774, + 0.007131520658731461, + 0.005253906361758709, + 0.003829976310953498, + 0.006353556644171476, + 0.003533766372129321, + 0.0027785729616880417, + 0.0030097984708845615, + 0.006634715478867292, + 0.005890561733394861, + 0.004751457832753658, + 0.005202620290219784, + 0.004326113034039736, + 0.003154907375574112, + 0.0023287988733500242, + 0.002928034169599414, + 0.004839365836232901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.003610759973526001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496004968881607 + ], + "opposite_logprob_tokens": [ + -2.929960012435913 + ], + "logprob_reward": -0.05496004968881607, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496004968881607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1934952437877655, + 0.10344485193490982, + 0.07054004073143005, + 0.1387975811958313, + 0.05039624869823456, + 0.17213262617588043, + 0.14329199492931366, + 0.03900996223092079, + 0.037046823650598526, + 0.051844604313373566 + ], + "credit_map": [ + 0.03150493651628494, + 0.04002196714282036, + 0.023287221789360046, + 0.025201376527547836, + 0.015316382981836796, + 0.014089900068938732, + 0.013651557266712189, + 0.030421895906329155, + 0.01884441077709198, + 0.016349265351891518, + 0.01726815663278103, + 0.014318459667265415, + 0.009374945424497128, + 0.009517473168671131, + 0.008748190477490425, + 0.009023955091834068, + 0.012576518580317497, + 0.009735256433486938, + 0.01330432016402483, + 0.010943840257823467, + 0.005957343615591526, + 0.004042000509798527, + 0.005657299421727657, + 0.008323460817337036, + 0.0259026400744915, + 0.03627576678991318, + 0.012802913784980774, + 0.01300984900444746, + 0.01744581013917923, + 0.01270798034965992, + 0.010423019528388977, + 0.010229597799479961, + 0.008643136359751225, + 0.008479548618197441, + 0.009077087976038456, + 0.005888555198907852, + 0.0035922066308557987, + 0.003300112672150135, + 0.0037037276197224855, + 0.007711871527135372, + 0.02622467651963234, + 0.014256752096116543, + 0.02674940600991249, + 0.03208718076348305, + 0.02289288118481636, + 0.009588513523340225, + 0.023750027641654015, + 0.016583183780312538, + 0.023429738357663155, + 0.014959340915083885, + 0.011656975373625755, + 0.01928933896124363, + 0.016015592962503433, + 0.014092466793954372, + 0.023732200264930725, + 0.02011633850634098, + 0.006012279074639082, + 0.005513746291399002, + 0.004050562623888254, + 0.007528737187385559, + 0.0047245873138308525, + 0.002838814863935113, + 0.003725075162947178, + 0.00461615901440382, + 0.006389673799276352, + 0.006861572153866291, + 0.007805989123880863, + 0.0047326842322945595, + 0.0029876611661165953, + 0.002056795870885253, + 0.0024264256935566664, + 0.0037860185839235783, + 0.009539508260786533, + 0.00796884298324585, + 0.004831768572330475, + 0.006608440540730953, + 0.005851647350937128, + 0.005590921733528376, + 0.005467824637889862, + 0.005985649302601814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.05496004968881607 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977255403995514 + ], + "opposite_logprob_tokens": [ + -4.634772777557373 + ], + "logprob_reward": -0.00977255403995514, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977255403995514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35805976390838623, + 0.12753631174564362, + 0.054588112980127335, + 0.09618775546550751, + 0.07513121515512466, + 0.0838843360543251, + 0.0738338753581047, + 0.05391160398721695, + 0.0436706468462944, + 0.033196404576301575 + ], + "credit_map": [ + 0.06260675936937332, + 0.09239169955253601, + 0.030101127922534943, + 0.0395481176674366, + 0.029908103868365288, + 0.02301444485783577, + 0.02823628857731819, + 0.05225323513150215, + 0.021677754819393158, + 0.021410148590803146, + 0.019246814772486687, + 0.017923220992088318, + 0.011648330837488174, + 0.012708134949207306, + 0.010777746327221394, + 0.012144160456955433, + 0.008108790963888168, + 0.005513942800462246, + 0.009414281696081161, + 0.009087695740163326, + 0.007171257399022579, + 0.003597414353862405, + 0.0073470487259328365, + 0.004347676876932383, + 0.018134748563170433, + 0.026048660278320312, + 0.01040758565068245, + 0.01032056100666523, + 0.007989551872015, + 0.0067332470789551735, + 0.0062058768235147, + 0.010347514413297176, + 0.014116093516349792, + 0.0160761047154665, + 0.007786641828715801, + 0.0076901293359696865, + 0.0071050808764994144, + 0.005087504629045725, + 0.006715747993439436, + 0.010553909465670586, + 0.012557990849018097, + 0.0060747829265892506, + 0.0134595837444067, + 0.014303219504654408, + 0.009060302749276161, + 0.004877803847193718, + 0.010715784505009651, + 0.01283486932516098, + 0.01165956910699606, + 0.007191869895905256, + 0.008356628008186817, + 0.008891143836081028, + 0.009359072893857956, + 0.011780640110373497, + 0.009221184067428112, + 0.007373768836259842, + 0.009572644717991352, + 0.01053935382515192, + 0.007993643172085285, + 0.009160349145531654, + 0.004041740205138922, + 0.0033690764103084803, + 0.0037714249920099974, + 0.005463374312967062, + 0.008201738819479942, + 0.0075274561531841755, + 0.005068269092589617, + 0.006460537668317556, + 0.0028114523738622665, + 0.002462267642840743, + 0.0032296767458319664, + 0.007909245789051056, + 0.005205114372074604, + 0.005627376958727837, + 0.0038745906203985214, + 0.0032734971027821302, + 0.004217149224132299, + 0.0025254038628190756, + 0.003100132802501321, + 0.005373138934373856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.00977255403995514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot at eye level captures a scene rendered in a vibrant, cheerful, cartoon-like comic book art style, defined by clean lines and bright, saturated colors. A young adult man with warm, olive skin and unkempt black hair is the focal point, dressed in a simple white T-shirt and a sky-blue apron. He leans slightly over a modern stainless steel stovetop, where a cherry-red enameled pot gently simmers its contents, a rich, warm orange. The entire kitchen is bathed in bright, even overhead lighting, making every color pop and eliminating harsh shadows. He carefully lifts a wooden tasting spoon from the pot and steadily brings it to his mouth, delicate wisps of steam curling upward. His expression is a perfect blend of concentration and delight—his brows furrowed slightly, but his eyes sparkle as he carefully savors the liquid, his lips curled in a subtle, satisfied smile, capturing the precise, joyful moment of perfecting his recipe in a warm and wholesome atmosphere.", + "seed": 23638, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_3272_seed_23638_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": -0.00805363985709846, + "vlm_logprob": -0.00805363985709846, + "vlm_yes_no_margin": 5.675, + "vlm_token_credit": -0.00805363985709846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0027903690934181213 + ], + "opposite_logprob_tokens": [ + -8.127790451049805 + ], + "logprob_reward": -0.0027903690934181213, + "yes_no_margin_reward": 8.125, + "token_credit_reward": -0.0027903690934181213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6455879211425781, + 0.06314001232385635, + 0.013052446767687798, + 0.1932288259267807, + 0.008752795867621899, + 0.007083195261657238, + 0.026158882305026054, + 0.009772655554115772, + 0.010012638755142689, + 0.023210611194372177 + ], + "credit_map": [ + 0.26616454124450684, + 0.06938167661428452, + 0.16633838415145874, + 0.04480157420039177, + 0.0275980606675148, + 0.04192693531513214, + 0.013278274796903133, + 0.01609843783080578, + 0.026401134207844734, + 0.00950813852250576, + 0.013952118344604969, + 0.004510978236794472, + 0.0029288632795214653, + 0.002321138745173812, + 0.0017546780873090029, + 0.0017629577778279781, + 0.002421274082735181, + 0.0015753961633890867, + 0.002712375484406948, + 0.001555806607939303, + 0.001260762452147901, + 0.001358836074359715, + 0.0010503462981432676, + 0.0011176486732438207, + 0.07651039958000183, + 0.022459974512457848, + 0.05668254196643829, + 0.012408825568854809, + 0.007189478259533644, + 0.00809804443269968, + 0.004338016267865896, + 0.005541552323848009, + 0.0018346511060371995, + 0.000992808025330305, + 0.001825795159675181, + 0.0009511248208582401, + 0.0007764965994283557, + 0.0010880863992497325, + 0.0007873661816120148, + 0.0004964680410921574, + 0.0012758569791913033, + 0.0009549890528433025, + 0.0007673780783079565, + 0.0009689368307590485, + 0.0008217656286433339, + 0.0010783860925585032, + 0.0006307324510999024, + 0.0005851494497619569, + 0.008630895055830479, + 0.002930321032181382, + 0.006176676135510206, + 0.0017513453494757414, + 0.0020072516053915024, + 0.002285028574988246, + 0.0010828305967152119, + 0.0012945343041792512, + 0.001667899894528091, + 0.0012218189658597112, + 0.001175421173684299, + 0.0008838546345941722, + 0.0012591852573677897, + 0.0017487232107669115, + 0.0009362003766000271, + 0.000879552448168397, + 0.0016586328856647015, + 0.0016688078176230192, + 0.0018568411469459534, + 0.0010716787073761225, + 0.0012261184165254235, + 0.0008829498547129333, + 0.0009331978508271277, + 0.0007144119590520859, + 0.004304084461182356, + 0.006733983755111694, + 0.003270562272518873, + 0.0021346292924135923, + 0.002323088003322482, + 0.00167503347620368, + 0.0012207619147375226, + 0.0015484672039747238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0027903690934181213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634116776287556 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004634116776287556, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634116776287556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5266692042350769, + 0.038478922098875046, + 0.025615399703383446, + 0.18947097659111023, + 0.01911938562989235, + 0.01697106659412384, + 0.0890376940369606, + 0.012171046808362007, + 0.010186463594436646, + 0.07227985560894012 + ], + "credit_map": [ + 0.2228100299835205, + 0.05768109858036041, + 0.13497291505336761, + 0.03630382567644119, + 0.020556356757879257, + 0.029164986684918404, + 0.011616752482950687, + 0.013563216663897038, + 0.011944272555410862, + 0.0061308518052101135, + 0.005913661327213049, + 0.003516184166073799, + 0.002683282596990466, + 0.0029051778838038445, + 0.0022950193379074335, + 0.0030904703307896852, + 0.006487211678177118, + 0.0035509965382516384, + 0.0023255464620888233, + 0.001749189686961472, + 0.0027434376534074545, + 0.00318387383595109, + 0.0023346063680946827, + 0.00324053387157619, + 0.07347302883863449, + 0.022161221131682396, + 0.05605384707450867, + 0.012022027745842934, + 0.00717510050162673, + 0.008223168551921844, + 0.004627855960279703, + 0.005734710022807121, + 0.006043585017323494, + 0.0019232736667618155, + 0.0028216291684657335, + 0.0012162324273958802, + 0.0017823620000854135, + 0.0026084696874022484, + 0.0013400384923443198, + 0.0013837921433150768, + 0.004677940160036087, + 0.001536283059976995, + 0.003021365962922573, + 0.0012275772169232368, + 0.0018388768658041954, + 0.002224701689556241, + 0.0010391053510829806, + 0.001405214425176382, + 0.033849991858005524, + 0.011143284849822521, + 0.02605060487985611, + 0.0051445914432406425, + 0.0032731574028730392, + 0.004220612812787294, + 0.002453017281368375, + 0.002902428386732936, + 0.0036942176520824432, + 0.001339531154371798, + 0.002128975698724389, + 0.0009693056344985962, + 0.001248852233402431, + 0.0012130162212997675, + 0.0008638220024295151, + 0.0007133249309845269, + 0.0028995652683079243, + 0.0011164873139932752, + 0.0018715703627094626, + 0.0009356269147247076, + 0.0008777478360570967, + 0.0010178120573982596, + 0.0007554600015282631, + 0.0007121920352801681, + 0.023953542113304138, + 0.010534217581152916, + 0.02114715985953808, + 0.004370494745671749, + 0.0033535733819007874, + 0.00423244945704937, + 0.0019380513112992048, + 0.002750356448814273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.004634116776287556 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019365858752280474 + ], + "opposite_logprob_tokens": [ + -6.251936435699463 + ], + "logprob_reward": -0.0019365858752280474, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019365858752280474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4492029547691345, + 0.02120477892458439, + 0.01565203070640564, + 0.365527480840683, + 0.015489738434553146, + 0.009698126465082169, + 0.022609340026974678, + 0.04384657368063927, + 0.007814078591763973, + 0.04895486310124397 + ], + "credit_map": [ + 0.17759042978286743, + 0.04681219533085823, + 0.11424761265516281, + 0.031954143196344376, + 0.02146628126502037, + 0.034579865634441376, + 0.010089506395161152, + 0.012462924234569073, + 0.006743758916854858, + 0.0020263788755983114, + 0.005031649488955736, + 0.0019654694478958845, + 0.001462658983655274, + 0.0015266294358298182, + 0.0013053863076493144, + 0.0011428477009758353, + 0.0031212265603244305, + 0.0016485315281897783, + 0.002566577633842826, + 0.001548625878058374, + 0.0019925138913094997, + 0.001983353868126869, + 0.0011946562444791198, + 0.0015965455677360296, + 0.14105837047100067, + 0.04279426485300064, + 0.1085100993514061, + 0.023975495249032974, + 0.014313955791294575, + 0.015947330743074417, + 0.008238550275564194, + 0.010689422488212585, + 0.0019509491976350546, + 0.0018461140571162105, + 0.0023663195315748453, + 0.0021325810812413692, + 0.0014512327034026384, + 0.0014866957208141685, + 0.002351434901356697, + 0.0019044120563194156, + 0.0018226888496428728, + 0.001334860222414136, + 0.0014165345346555114, + 0.001300877658650279, + 0.001214821357280016, + 0.0009162928909063339, + 0.001119838678278029, + 0.0005722123314626515, + 0.008769039995968342, + 0.002896985737606883, + 0.005242418963462114, + 0.0012848565820604563, + 0.00128414505161345, + 0.001273494679480791, + 0.0010737036354839802, + 0.0007846941007301211, + 0.007087607402354479, + 0.00411048112437129, + 0.0054237125441432, + 0.0046678343787789345, + 0.005203273147344589, + 0.005104144103825092, + 0.008055059239268303, + 0.0041944594122469425, + 0.0010816652793437243, + 0.001024206867441535, + 0.0016522102523595095, + 0.0011932812631130219, + 0.000685108476318419, + 0.0005765625392086804, + 0.0010533684398978949, + 0.0005476756487041712, + 0.01556622888892889, + 0.006697376724332571, + 0.013466283679008484, + 0.0032497637439519167, + 0.002862465800717473, + 0.003326523583382368, + 0.001730293850414455, + 0.0020559297408908606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0019365858752280474 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.36629292368888855, + 0.03019271418452263, + 0.029345238581299782, + 0.21305184066295624, + 0.02679973654448986, + 0.013191617093980312, + 0.16767998039722443, + 0.07732832431793213, + 0.008592900820076466, + 0.06752469390630722 + ], + "credit_map": [ + 0.14946268498897552, + 0.038994692265987396, + 0.09400675445795059, + 0.025476133450865746, + 0.01616796664893627, + 0.02481914684176445, + 0.007786481641232967, + 0.009579078294336796, + 0.010860350914299488, + 0.004899327177554369, + 0.004833064042031765, + 0.0029932258184999228, + 0.0016947024269029498, + 0.001442824024707079, + 0.0021139662712812424, + 0.0013552518794313073, + 0.010379880666732788, + 0.0038605965673923492, + 0.006302970927208662, + 0.00246642529964447, + 0.0017814935417845845, + 0.001512401388026774, + 0.0013752715894952416, + 0.0016661996487528086, + 0.08214090019464493, + 0.02437250129878521, + 0.0616346150636673, + 0.014077805913984776, + 0.008694002404808998, + 0.010032383725047112, + 0.005093191750347614, + 0.007006450090557337, + 0.007628395222127438, + 0.0028458519373089075, + 0.0037458729930222034, + 0.002027606824412942, + 0.0028428155928850174, + 0.0040350304916501045, + 0.0018981691682711244, + 0.0017759939655661583, + 0.0020446190610527992, + 0.001565456041134894, + 0.0016307767946273088, + 0.0011717062443494797, + 0.001999161671847105, + 0.0016620620153844357, + 0.0017622550949454308, + 0.0013555800542235374, + 0.06377612799406052, + 0.02137969620525837, + 0.04928005859255791, + 0.009247650392353535, + 0.0060577052645385265, + 0.008409349247813225, + 0.004397382494062185, + 0.005131993442773819, + 0.020902257412672043, + 0.00766452681273222, + 0.008357306011021137, + 0.007039449643343687, + 0.008136129006743431, + 0.008637604303658009, + 0.010351558215916157, + 0.006239485926926136, + 0.001455176854506135, + 0.0009188276017084718, + 0.0013951598666608334, + 0.001049019512720406, + 0.0009407639154233038, + 0.0010059436317533255, + 0.0011628876673057675, + 0.0006651227013207972, + 0.024205327033996582, + 0.008398591540753841, + 0.019358811900019646, + 0.004117689095437527, + 0.003249360015615821, + 0.003540370613336563, + 0.0021273167803883553, + 0.0025272287894040346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.02328539453446865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621733006089926 + ], + "opposite_logprob_tokens": [ + -4.882621765136719 + ], + "logprob_reward": -0.007621733006089926, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621733006089926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2455824911594391, + 0.15084078907966614, + 0.0629555881023407, + 0.20059135556221008, + 0.03280424326658249, + 0.01410349179059267, + 0.13147228956222534, + 0.02341962419450283, + 0.019466230645775795, + 0.11876396089792252 + ], + "credit_map": [ + 0.09176087379455566, + 0.025582751259207726, + 0.05938660353422165, + 0.017912443727254868, + 0.014995009638369083, + 0.019690632820129395, + 0.008558787405490875, + 0.0076953573152422905, + 0.06534641236066818, + 0.025092637166380882, + 0.029867086559534073, + 0.011132639832794666, + 0.006392713636159897, + 0.004480478819459677, + 0.003804996842518449, + 0.004723789170384407, + 0.00921669416129589, + 0.016347769647836685, + 0.016979198902845383, + 0.0049847871996462345, + 0.0045686084777116776, + 0.004500380251556635, + 0.0032525232527405024, + 0.0031056194566190243, + 0.07521522045135498, + 0.023700051009655, + 0.05948013812303543, + 0.013363391160964966, + 0.008397459052503109, + 0.009519997052848339, + 0.004909694194793701, + 0.0060053798370063305, + 0.009360276162624359, + 0.002954284194856882, + 0.006081767845898867, + 0.002811908023431897, + 0.0023329704999923706, + 0.0037665609270334244, + 0.0028564909007400274, + 0.0026399819180369377, + 0.002301872009411454, + 0.002062877407297492, + 0.0011135797249153256, + 0.0015263463137671351, + 0.0022168292198330164, + 0.0019979646895080805, + 0.0013347999192774296, + 0.0015492219245061278, + 0.0508706197142601, + 0.01640581339597702, + 0.03767331689596176, + 0.0072554899379611015, + 0.0045568011701107025, + 0.0065594688057899475, + 0.004075747448951006, + 0.004075019620358944, + 0.0037702955305576324, + 0.0029340856708586216, + 0.004327036906033754, + 0.00342767802067101, + 0.0028797283302992582, + 0.001974145881831646, + 0.0020846629049628973, + 0.002021987223997712, + 0.005291771609336138, + 0.002135114511474967, + 0.004904064815491438, + 0.0017350007547065616, + 0.0013222963316366076, + 0.001481315353885293, + 0.001583568868227303, + 0.0010130970040336251, + 0.04033760726451874, + 0.017573492601513863, + 0.03632858395576477, + 0.0073322393000125885, + 0.005041445139795542, + 0.0053987749852240086, + 0.003107819240540266, + 0.003643979784101248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007621733006089926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot at eye level captures a scene rendered in a vibrant, cheerful, cartoon-like comic book art style, defined by clean lines and bright, saturated colors. A young adult man with warm, olive skin and unkempt black hair is the focal point, dressed in a simple white T-shirt and a sky-blue apron. He leans slightly over a modern stainless steel stovetop, where a cherry-red enameled pot gently simmers its contents, a rich, warm orange. The entire kitchen is bathed in bright, even overhead lighting, making every color pop and eliminating harsh shadows. He carefully lifts a wooden tasting spoon from the pot and steadily brings it to his mouth, delicate wisps of steam curling upward. His expression is a perfect blend of concentration and delight—his brows furrowed slightly, but his eyes sparkle as he carefully savors the liquid, his lips curled in a subtle, satisfied smile, capturing the precise, joyful moment of perfecting his recipe in a warm and wholesome atmosphere.", + "seed": 23639, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_3272_seed_23639_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": -0.00622504367493093, + "vlm_logprob": -0.00622504367493093, + "vlm_yes_no_margin": 6.2, + "vlm_token_credit": -0.00622504367493093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0053177098743617535 + ], + "opposite_logprob_tokens": [ + -9.630317687988281 + ], + "logprob_reward": -0.0053177098743617535, + "yes_no_margin_reward": 9.625, + "token_credit_reward": -0.0053177098743617535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10684489458799362, + 0.35802730917930603, + 0.05076037719845772, + 0.1333545297384262, + 0.06429184973239899, + 0.055138714611530304, + 0.04567690193653107, + 0.05282728001475334, + 0.09318345040082932, + 0.03989467769861221 + ], + "credit_map": [ + 0.013621969148516655, + 0.012591064907610416, + 0.018481725826859474, + 0.016511719673871994, + 0.015012339688837528, + 0.007581728044897318, + 0.01581469736993313, + 0.007229648530483246, + 0.1394410878419876, + 0.041910570114851, + 0.033114876598119736, + 0.025704508647322655, + 0.04175274819135666, + 0.015553773380815983, + 0.012667869217693806, + 0.04788189008831978, + 0.009943010285496712, + 0.009462106972932816, + 0.007832305505871773, + 0.006231199484318495, + 0.0065126437693834305, + 0.0024930357467383146, + 0.005469369702041149, + 0.0028167073614895344, + 0.021947775036096573, + 0.017461242154240608, + 0.040056757628917694, + 0.010147647932171822, + 0.019789382815361023, + 0.007145934272557497, + 0.010182646103203297, + 0.006623149383813143, + 0.010774602182209492, + 0.01004735380411148, + 0.012549785897135735, + 0.011740158312022686, + 0.007177135907113552, + 0.0033621778711676598, + 0.005143747664988041, + 0.003496886696666479, + 0.007520819548517466, + 0.009548609144985676, + 0.0052961185574531555, + 0.01233960222452879, + 0.004595509264618158, + 0.003101183334365487, + 0.007380039431154728, + 0.005356830079108477, + 0.006413828115910292, + 0.005161552224308252, + 0.007434063591063023, + 0.00975428894162178, + 0.005720219574868679, + 0.003408754011616111, + 0.004274224396795034, + 0.003509970149025321, + 0.0076472158543765545, + 0.006234190426766872, + 0.008327714167535305, + 0.015745462849736214, + 0.005343322176486254, + 0.0029692540410906076, + 0.004528867546468973, + 0.002031253185123205, + 0.013805519789457321, + 0.00931003037840128, + 0.019273491576313972, + 0.02232031710445881, + 0.00923079438507557, + 0.004701914731413126, + 0.009906797669827938, + 0.004634590353816748, + 0.006693956907838583, + 0.005003598518669605, + 0.006239360198378563, + 0.007754408288747072, + 0.004543793387711048, + 0.0033607410732656717, + 0.0040510729886591434, + 0.0022477498278021812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0053177098743617535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634377270936966 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634377270936966, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634377270936966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.156450554728508, + 0.2867436408996582, + 0.17202305793762207, + 0.09107261151075363, + 0.03304849565029144, + 0.07000067830085754, + 0.0724189504981041, + 0.036227717995643616, + 0.04390369728207588, + 0.038110606372356415 + ], + "credit_map": [ + 0.027590475976467133, + 0.01482298132032156, + 0.027937671169638634, + 0.014143100939691067, + 0.019396407529711723, + 0.011852066963911057, + 0.020280746743083, + 0.020427117124199867, + 0.10614815354347229, + 0.03320867568254471, + 0.026541676372289658, + 0.020891789346933365, + 0.03391429781913757, + 0.013185309246182442, + 0.014993026852607727, + 0.03786073252558708, + 0.03153914585709572, + 0.029132945463061333, + 0.025228722020983696, + 0.015963854268193245, + 0.024820541962981224, + 0.018065650016069412, + 0.013438187539577484, + 0.013834018260240555, + 0.01696590520441532, + 0.013326043263077736, + 0.03137011453509331, + 0.007093886844813824, + 0.009363334625959396, + 0.004466953221708536, + 0.004771799314767122, + 0.003714572638273239, + 0.006553334183990955, + 0.006455144379287958, + 0.005762659013271332, + 0.00438681710511446, + 0.0038794842548668385, + 0.0019033125136047602, + 0.002170054940506816, + 0.0019376875134184957, + 0.011144503951072693, + 0.010975970886647701, + 0.0065252468921244144, + 0.01980859786272049, + 0.006272295489907265, + 0.0033505847677588463, + 0.008195486851036549, + 0.003727993695065379, + 0.0104181207716465, + 0.011320752091705799, + 0.011384272947907448, + 0.012546684592962265, + 0.009308620356023312, + 0.004603156819939613, + 0.007250841706991196, + 0.005586504470556974, + 0.006181939039379358, + 0.0032969259191304445, + 0.007772058714181185, + 0.006623440887778997, + 0.005179329309612513, + 0.002012884244322777, + 0.003381650662049651, + 0.0017794902669265866, + 0.007575425319373608, + 0.00481092045083642, + 0.009480607695877552, + 0.006210354156792164, + 0.0048736766912043095, + 0.0036171115934848785, + 0.0050823744386434555, + 0.0022532299626618624, + 0.006592809222638607, + 0.009022025391459465, + 0.003970704507082701, + 0.004333149176090956, + 0.004493333399295807, + 0.003974591847509146, + 0.002609723014757037, + 0.00311427004635334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.008634377270936966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009182051289826632 + ], + "opposite_logprob_tokens": [ + -7.000918388366699 + ], + "logprob_reward": -0.0009182051289826632, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009182051289826632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08283869177103043, + 0.2726680636405945, + 0.1020548865199089, + 0.040164291858673096, + 0.0348471999168396, + 0.04789257422089577, + 0.04844556748867035, + 0.06981883198022842, + 0.26215383410453796, + 0.03911606967449188 + ], + "credit_map": [ + 0.01179569959640503, + 0.01014302484691143, + 0.01711895316839218, + 0.010876676067709923, + 0.010453801602125168, + 0.006039894185960293, + 0.009099222719669342, + 0.007311423774808645, + 0.10615663230419159, + 0.03319884091615677, + 0.027226341888308525, + 0.018041273579001427, + 0.031458575278520584, + 0.012021954171359539, + 0.009904992766678333, + 0.034659452736377716, + 0.018678652122616768, + 0.02338075451552868, + 0.020763462409377098, + 0.013229122385382652, + 0.008524799719452858, + 0.004940600600093603, + 0.008050587959587574, + 0.004486905410885811, + 0.010674094781279564, + 0.005116110667586327, + 0.014384939335286617, + 0.0027842908166348934, + 0.002150741871446371, + 0.0019412415567785501, + 0.001514130039140582, + 0.0015987419756129384, + 0.004634623881429434, + 0.004136217758059502, + 0.00489187752828002, + 0.009453055448830128, + 0.003794744610786438, + 0.0022741525899618864, + 0.003191923024132848, + 0.0024706032127141953, + 0.007230584975332022, + 0.010986967012286186, + 0.006636934354901314, + 0.00803002156317234, + 0.004877367056906223, + 0.0028272473718971014, + 0.004461158532649279, + 0.0028422935865819454, + 0.007094685919582844, + 0.005002365913242102, + 0.008352743461728096, + 0.011836020275950432, + 0.005571151617914438, + 0.0025894150603562593, + 0.004157769028097391, + 0.003841415047645569, + 0.013304991647601128, + 0.005420406814664602, + 0.020270735025405884, + 0.01306425966322422, + 0.00606164475902915, + 0.002967622596770525, + 0.006118803285062313, + 0.0026103665586560965, + 0.054823655635118484, + 0.017912065610289574, + 0.08023826032876968, + 0.042089954018592834, + 0.021597500890493393, + 0.010971799492835999, + 0.0249561108648777, + 0.009564490057528019, + 0.006072305142879486, + 0.005285108461976051, + 0.0056776911951601505, + 0.009408564306795597, + 0.003755858400836587, + 0.0029265200719237328, + 0.0031673419289290905, + 0.0028226773720234632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.0009182051289826632 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634612895548344 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634612895548344, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634612895548344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22546446323394775, + 0.24556180834770203, + 0.05879459157586098, + 0.04926964268088341, + 0.18416140973567963, + 0.05531063675880432, + 0.05009732395410538, + 0.049242060631513596, + 0.04487792029976845, + 0.03722013533115387 + ], + "credit_map": [ + 0.033141396939754486, + 0.025941135361790657, + 0.022680046036839485, + 0.040123432874679565, + 0.04448501765727997, + 0.024600081145763397, + 0.020404133945703506, + 0.014089218340814114, + 0.0920364037156105, + 0.02836150862276554, + 0.027563266456127167, + 0.017297841608524323, + 0.0300710778683424, + 0.011714213527739048, + 0.008823039941489697, + 0.029694460332393646, + 0.012036357074975967, + 0.010140901431441307, + 0.007483285386115313, + 0.00716440100222826, + 0.006725614424794912, + 0.003424584399908781, + 0.007109091151505709, + 0.004710356239229441, + 0.012544920668005943, + 0.0057237413711845875, + 0.014707420021295547, + 0.003935026470571756, + 0.004474346060305834, + 0.0023016987834125757, + 0.0033905114978551865, + 0.002191974548622966, + 0.0475570447742939, + 0.015458386391401291, + 0.04297928512096405, + 0.03454555943608284, + 0.01572439819574356, + 0.00757587980479002, + 0.012404307723045349, + 0.007916538044810295, + 0.006607686169445515, + 0.006509122438728809, + 0.009705434553325176, + 0.013722576200962067, + 0.005393250845372677, + 0.002935796044766903, + 0.006186102516949177, + 0.004250669851899147, + 0.007608000189065933, + 0.0076177679002285, + 0.009794572368264198, + 0.010612414218485355, + 0.0044187153689563274, + 0.002805972471833229, + 0.003981156274676323, + 0.003258724696934223, + 0.00836501456797123, + 0.004164526239037514, + 0.013517088256776333, + 0.011858242563903332, + 0.003223129315301776, + 0.0019294851226732135, + 0.0034098217729479074, + 0.0027747531421482563, + 0.007178901229053736, + 0.004584007430821657, + 0.012991600669920444, + 0.00892805214971304, + 0.00361158256419003, + 0.0017986646853387356, + 0.0038967328146100044, + 0.0018883822485804558, + 0.006188357714563608, + 0.004046540707349777, + 0.006116947159171104, + 0.008852813392877579, + 0.003477325662970543, + 0.002853844780474901, + 0.0036250357516109943, + 0.0020592703949660063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.008634612895548344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076203132048249245 + ], + "opposite_logprob_tokens": [ + -4.882620334625244 + ], + "logprob_reward": -0.0076203132048249245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076203132048249245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17158143222332, + 0.4427722096443176, + 0.06389933079481125, + 0.06707219779491425, + 0.09703230857849121, + 0.04300769418478012, + 0.03333865851163864, + 0.0358910858631134, + 0.02485724724829197, + 0.020547879859805107 + ], + "credit_map": [ + 0.025274313986301422, + 0.02698986604809761, + 0.02285466156899929, + 0.016604775562882423, + 0.024957703426480293, + 0.017403138801455498, + 0.01441913191229105, + 0.023077839985489845, + 0.1718655377626419, + 0.05513838678598404, + 0.041714902967214584, + 0.03138677403330803, + 0.05278291180729866, + 0.019974568858742714, + 0.014084401540458202, + 0.055824704468250275, + 0.013689523562788963, + 0.01330272201448679, + 0.010672541335225105, + 0.006163592915982008, + 0.008003165014088154, + 0.0040548453107476234, + 0.00402050418779254, + 0.003992436919361353, + 0.011461615562438965, + 0.01277115661650896, + 0.014444621279835701, + 0.0049138483591377735, + 0.00939076766371727, + 0.004115926567465067, + 0.005072001367807388, + 0.004902257118374109, + 0.028211675584316254, + 0.0081827687099576, + 0.03313428908586502, + 0.010516063310205936, + 0.004413491114974022, + 0.003134680911898613, + 0.004491779953241348, + 0.004947565495967865, + 0.006011148449033499, + 0.00996571034193039, + 0.005166369490325451, + 0.005328240804374218, + 0.004731645807623863, + 0.0031563108786940575, + 0.004817663691937923, + 0.003830604488030076, + 0.004354259464889765, + 0.004052306059747934, + 0.0047353943809866905, + 0.008355546742677689, + 0.0035012068692594767, + 0.0018684881506487727, + 0.003673114348202944, + 0.0027983388863503933, + 0.005114059895277023, + 0.004044204950332642, + 0.005794647615402937, + 0.009142288006842136, + 0.003672268707305193, + 0.0020416050683707, + 0.0038686704356223345, + 0.002213337691500783, + 0.004740957636386156, + 0.00200088182464242, + 0.007642755284905434, + 0.0028601964004337788, + 0.0024766121059656143, + 0.001150645548477769, + 0.0025476678274571896, + 0.0014375289902091026, + 0.0027668229304254055, + 0.0029416594188660383, + 0.0029126231092959642, + 0.0037911233957856894, + 0.0023340063635259867, + 0.001857540919445455, + 0.0024416535161435604, + 0.0015024479944258928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0076203132048249245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot at eye level captures a scene rendered in a vibrant, cheerful, cartoon-like comic book art style, defined by clean lines and bright, saturated colors. A young adult man with warm, olive skin and unkempt black hair is the focal point, dressed in a simple white T-shirt and a sky-blue apron. He leans slightly over a modern stainless steel stovetop, where a cherry-red enameled pot gently simmers its contents, a rich, warm orange. The entire kitchen is bathed in bright, even overhead lighting, making every color pop and eliminating harsh shadows. He carefully lifts a wooden tasting spoon from the pot and steadily brings it to his mouth, delicate wisps of steam curling upward. His expression is a perfect blend of concentration and delight—his brows furrowed slightly, but his eyes sparkle as he carefully savors the liquid, his lips curled in a subtle, satisfied smile, capturing the precise, joyful moment of perfecting his recipe in a warm and wholesome atmosphere.", + "seed": 23640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_2/index_3272_seed_23640_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": -0.005728514259681105, + "vlm_logprob": -0.005728514259681105, + "vlm_yes_no_margin": 5.9, + "vlm_token_credit": -0.005728514259681105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024034681264311075 + ], + "opposite_logprob_tokens": [ + -8.502403259277344 + ], + "logprob_reward": -0.0024034681264311075, + "yes_no_margin_reward": 8.5, + "token_credit_reward": -0.0024034681264311075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10852178931236267, + 0.03785713016986847, + 0.1343238800764084, + 0.09649962186813354, + 0.06588272005319595, + 0.04207129776477814, + 0.062295202165842056, + 0.04004949331283569, + 0.06256761401891708, + 0.3499312102794647 + ], + "credit_map": [ + 0.02074938453733921, + 0.011458742432296276, + 0.016392188146710396, + 0.013962343335151672, + 0.008207490667700768, + 0.011354430578649044, + 0.013596509583294392, + 0.012800702825188637, + 0.006858513690531254, + 0.003852967405691743, + 0.007600109558552504, + 0.004063545260578394, + 0.002863784320652485, + 0.003833367954939604, + 0.005653713829815388, + 0.0031311283819377422, + 0.049625542014837265, + 0.0065667214803397655, + 0.036789651960134506, + 0.008602018468081951, + 0.007698864210397005, + 0.0075385477393865585, + 0.011235534213483334, + 0.006267012562602758, + 0.016473636031150818, + 0.013636820949614048, + 0.012072729878127575, + 0.009699413552880287, + 0.007724026683717966, + 0.014992299489676952, + 0.008989997208118439, + 0.012910706922411919, + 0.012305142357945442, + 0.0076215569861233234, + 0.008573236875236034, + 0.0060360245406627655, + 0.007739916443824768, + 0.009492110460996628, + 0.007175936363637447, + 0.0069387988187372684, + 0.009812939912080765, + 0.004256682004779577, + 0.00391326704993844, + 0.004656478296965361, + 0.00449620746076107, + 0.006325891707092524, + 0.004494026303291321, + 0.004115805961191654, + 0.010647800751030445, + 0.005299973767250776, + 0.008009666576981544, + 0.005404318682849407, + 0.00608325470238924, + 0.0057176630944013596, + 0.014998644590377808, + 0.006133882328867912, + 0.005826898850500584, + 0.0029589852783828974, + 0.006002354901283979, + 0.003221081104129553, + 0.0034004943445324898, + 0.0032901298254728317, + 0.011489592492580414, + 0.0038599553517997265, + 0.015475206077098846, + 0.0056244079023599625, + 0.010060516186058521, + 0.005080964881926775, + 0.00509735057130456, + 0.010371636599302292, + 0.006002705078572035, + 0.0048548271879553795, + 0.09440569579601288, + 0.018437296152114868, + 0.028297241777181625, + 0.04609311372041702, + 0.019721070304512978, + 0.021378343924880028, + 0.091712087392807, + 0.029886338859796524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.0024034681264311075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486948622390628 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.002486948622390628, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486948622390628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2101895958185196, + 0.04487655311822891, + 0.17883704602718353, + 0.11780724674463272, + 0.07703669369220734, + 0.05080750212073326, + 0.04052216187119484, + 0.08570419996976852, + 0.03492685779929161, + 0.15929214656352997 + ], + "credit_map": [ + 0.04176560044288635, + 0.034013669937849045, + 0.02830956131219864, + 0.027259165421128273, + 0.01830221712589264, + 0.028296759352087975, + 0.01932809129357338, + 0.012914503924548626, + 0.007474364712834358, + 0.0032341410405933857, + 0.00630990881472826, + 0.005363142117857933, + 0.003408373100683093, + 0.004046954680234194, + 0.011019992642104626, + 0.0040196687914431095, + 0.06838414818048477, + 0.007958497852087021, + 0.050669871270656586, + 0.012168881483376026, + 0.006347217131406069, + 0.010309480130672455, + 0.013633385300636292, + 0.009365539066493511, + 0.025472557172179222, + 0.01137875858694315, + 0.016233276575803757, + 0.011144670657813549, + 0.008896232582628727, + 0.017248015850782394, + 0.016279814764857292, + 0.011153903789818287, + 0.013525044545531273, + 0.011114981025457382, + 0.009350696578621864, + 0.007278506178408861, + 0.009144928306341171, + 0.010786653496325016, + 0.008488234132528305, + 0.007347642909735441, + 0.008877091109752655, + 0.004453182220458984, + 0.009532222524285316, + 0.004682451486587524, + 0.004553730133920908, + 0.004635953810065985, + 0.007522071246057749, + 0.006550794001668692, + 0.0073813823983073235, + 0.004440498072654009, + 0.006156545598059893, + 0.0031671952456235886, + 0.0050918906927108765, + 0.005178789608180523, + 0.004855493549257517, + 0.004250365309417248, + 0.012511005625128746, + 0.006045853719115257, + 0.013783962465822697, + 0.005862635560333729, + 0.0075560640543699265, + 0.006949699018150568, + 0.025698600336909294, + 0.007296363357454538, + 0.006494054105132818, + 0.0028096947353333235, + 0.006660921033471823, + 0.0035208489280194044, + 0.003097702981904149, + 0.0034220656380057335, + 0.004794182721525431, + 0.004127384629100561, + 0.0354095958173275, + 0.00940747745335102, + 0.016392609104514122, + 0.02119525521993637, + 0.010958392173051834, + 0.012995594181120396, + 0.03989366441965103, + 0.013039548881351948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.002486948622390628 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613254288211465 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613254288211465, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613254288211465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13104863464832306, + 0.043105363845825195, + 0.1342930644750595, + 0.03193339332938194, + 0.07070417702198029, + 0.02801392413675785, + 0.03419854864478111, + 0.02046009711921215, + 0.0417185053229332, + 0.4645242691040039 + ], + "credit_map": [ + 0.025608109310269356, + 0.019474143162369728, + 0.0184027086943388, + 0.018530840054154396, + 0.009297631680965424, + 0.015708694234490395, + 0.014596760272979736, + 0.009429743513464928, + 0.00881691463291645, + 0.002658693352714181, + 0.00907941535115242, + 0.006923097185790539, + 0.0025375611148774624, + 0.0037108492106199265, + 0.00526772765442729, + 0.004111102782189846, + 0.04620790854096413, + 0.006291297730058432, + 0.036024875938892365, + 0.01231847982853651, + 0.005646756384521723, + 0.008146915584802628, + 0.01304531004279852, + 0.0066115278750658035, + 0.004408418200910091, + 0.004630906041711569, + 0.0044042449444532394, + 0.004221890587359667, + 0.002791221486404538, + 0.005676810164004564, + 0.0029168555047363043, + 0.002883046166971326, + 0.011237739585340023, + 0.007720980793237686, + 0.0069282036274671555, + 0.007484201807528734, + 0.010024120099842548, + 0.011829114519059658, + 0.00851436983793974, + 0.0069654411636292934, + 0.007112765219062567, + 0.0020561194978654385, + 0.002704343991354108, + 0.0025612530298531055, + 0.002545930678024888, + 0.0037850658409297466, + 0.0037364973686635494, + 0.003511949209496379, + 0.009518398903310299, + 0.0029051636811345816, + 0.003239578101783991, + 0.0026601082645356655, + 0.0024619237519800663, + 0.002818392589688301, + 0.006956912577152252, + 0.0036380710080266, + 0.005136388354003429, + 0.0018334246706217527, + 0.0027375142090022564, + 0.0014358473708853126, + 0.001602801843546331, + 0.0016741829458624125, + 0.003620703238993883, + 0.0024192340206354856, + 0.0061754803173244, + 0.0037937783636152744, + 0.004408963490277529, + 0.002988502848893404, + 0.003776314202696085, + 0.004236745182424784, + 0.012326580472290516, + 0.0040121437050402164, + 0.12244401127099991, + 0.024646909907460213, + 0.03957350552082062, + 0.06569280475378036, + 0.027176091447472572, + 0.02600627951323986, + 0.11682398617267609, + 0.04216067120432854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.003613254288211465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195299707353115 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195299707353115, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195299707353115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032854944467544556, + 0.010877023451030254, + 0.044535111635923386, + 0.02493223547935486, + 0.018062861636281013, + 0.018771909177303314, + 0.018316451460123062, + 0.031792961061000824, + 0.035338692367076874, + 0.7645177841186523 + ], + "credit_map": [ + 0.006079306825995445, + 0.004085173364728689, + 0.005056804046034813, + 0.004215392284095287, + 0.002454145345836878, + 0.004062286112457514, + 0.004045385867357254, + 0.002856456208974123, + 0.0024691985454410315, + 0.000749590341001749, + 0.0015985865611582994, + 0.0009567262604832649, + 0.0006530672544613481, + 0.001234554685652256, + 0.0010700360871851444, + 0.002145265694707632, + 0.014096196740865707, + 0.0019086883403360844, + 0.01210850104689598, + 0.005416736472398043, + 0.0018104788614436984, + 0.0024659179616719484, + 0.003908854443579912, + 0.0028197444044053555, + 0.0045656743459403515, + 0.0030910277273505926, + 0.004418539814651012, + 0.0034864516928792, + 0.0016101215733215213, + 0.002241627313196659, + 0.0025709623005241156, + 0.0029478371143341064, + 0.003300786716863513, + 0.001750481897033751, + 0.0020861492957919836, + 0.002447988837957382, + 0.002039761980995536, + 0.001781064085662365, + 0.002984667895361781, + 0.0016719632549211383, + 0.004125053063035011, + 0.0012950592208653688, + 0.002036014571785927, + 0.001919603324495256, + 0.00146797439083457, + 0.003133748658001423, + 0.0026557734236121178, + 0.0021386859007179737, + 0.0027804807759821415, + 0.0018274863250553608, + 0.0031737172976136208, + 0.001742078340612352, + 0.0013997963396832347, + 0.0032482733950018883, + 0.0019769538193941116, + 0.0021676686592400074, + 0.004459825344383717, + 0.0019935728050768375, + 0.0059641702100634575, + 0.003774499287828803, + 0.0023978338576853275, + 0.0024950343649834394, + 0.007838954217731953, + 0.002869074232876301, + 0.007844016887247562, + 0.0036833561025559902, + 0.0036445993464440107, + 0.0020444432739168406, + 0.0029165097512304783, + 0.007690484635531902, + 0.003900169162079692, + 0.0036151213571429253, + 0.20619122684001923, + 0.040071744471788406, + 0.06335008144378662, + 0.10714881122112274, + 0.0451100654900074, + 0.04103398323059082, + 0.191119983792305, + 0.0704919695854187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.014195299707353115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a vibrant, cheerful, cartoon-like comic book art style with clean lines and bright, saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point a young adult man with warm, olive skin, unkempt black hair, wearing a white T-shirt and a sky-blue apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man leaning over a modern stainless steel stovetop with a cherry-red enameled pot containing a rich, warm orange liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man lift a wooden tasting spoon from the pot and bring it to his mouth while steam curls upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's expression showing concentration and delight with furrowed brows, sparkling eyes, and a subtle, satisfied smile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943600554019213 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943600554019213, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943600554019213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13181613385677338, + 0.04226524382829666, + 0.12977755069732666, + 0.0960756167769432, + 0.06071213632822037, + 0.09666360914707184, + 0.06791912019252777, + 0.09385750442743301, + 0.09001263231039047, + 0.19090048968791962 + ], + "credit_map": [ + 0.025271806865930557, + 0.021552303805947304, + 0.01600557006895542, + 0.014492236077785492, + 0.01057988777756691, + 0.01831085979938507, + 0.01634727232158184, + 0.009256178513169289, + 0.00594262545928359, + 0.0037852919194847345, + 0.006329541560262442, + 0.004703105427324772, + 0.003984817769378424, + 0.005096494685858488, + 0.005204328335821629, + 0.007219033781439066, + 0.04880489036440849, + 0.005512296222150326, + 0.03876704350113869, + 0.008356091566383839, + 0.004310260061174631, + 0.007289309520274401, + 0.010113198310136795, + 0.0066244644112885, + 0.0222754068672657, + 0.011544212698936462, + 0.011219128966331482, + 0.009165394119918346, + 0.008222999051213264, + 0.013545316644012928, + 0.009757908061146736, + 0.010345234535634518, + 0.009125806391239166, + 0.008406068198382854, + 0.005228503607213497, + 0.005339737515896559, + 0.007494647055864334, + 0.009389295242726803, + 0.0074789379723370075, + 0.008249138481914997, + 0.01962020993232727, + 0.009182358160614967, + 0.011409332975745201, + 0.008384490385651588, + 0.010615240782499313, + 0.010048521682620049, + 0.013068613596260548, + 0.014334836974740028, + 0.011899773962795734, + 0.0039858766831457615, + 0.00922976154834032, + 0.00611617136746645, + 0.005454426631331444, + 0.004939555190503597, + 0.018764512613415718, + 0.007529041729867458, + 0.018926892429590225, + 0.005399356130510569, + 0.014331096783280373, + 0.008301664143800735, + 0.006759196054190397, + 0.004539787769317627, + 0.027881555259227753, + 0.007717945147305727, + 0.018701741471886635, + 0.007370466366410255, + 0.014813896268606186, + 0.00644885515794158, + 0.006267212796956301, + 0.013846147805452347, + 0.014673760160803795, + 0.007890540175139904, + 0.04188867658376694, + 0.01009557768702507, + 0.01617281511425972, + 0.03314671292901039, + 0.01588231511414051, + 0.012311652302742004, + 0.03202711418271065, + 0.02937561646103859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.005943600554019213 + } + ], + "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_3/index_tensor([7963, 7963])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_tensor([7963, 7963])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..60cb54731e59141e2df3d60a21a274ddea066944 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_tensor([7963, 7963])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The hyperreal 3D rendering begins with a low-angle, uninterrupted, doll-like shot, gliding smoothly and silently forward across the polished, reflective white epoxy floor of a vast, sterile semiconductor manufacturing facility. A fine, clean mist wafts through the cool, filtered air, capturing and diffusing the clinical glow of the embedded LED panels overhead and the vibrant blues and ambers of status lights on the towering, complex machinery. The camera's path gracefully curves toward a multi-jointed robotic arm, its polished chrome and white composite surfaces gleaming as it extends with silent, deliberate precision. A specialized vacuum gripper at its end delicately removes a gleaming, perfectly round silicon chip from its box, its surface brimming with iridescent microscopic circuitry, refracting light into a rainbow of colors. As the robotic arm rotates and places the chip into a processing chamber, the chip's fluid journey complete. The camera pans left to reveal a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield. They stand in front of a bank of monitors with focused precision. As complex streams of data scroll across the screens, their gloved fingers methodically interact with the control interfaces, their gestures conveying absolute concentration in the facility's hushed surroundings. The camera continues its seamless forward movement, deep into the high-tech maze where human focus and automated perfection operate in a silent, synchronized ballet, creating an atmosphere of profound, futuristic tranquility and meticulous creation.", + "seed": 6343, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_7963_seed_6343_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": -0.0405373964458704, + "vlm_logprob": -0.0405373964458704, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.0405373964458704, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013897789642214775 + ], + "opposite_logprob_tokens": [ + -5.263897895812988 + ], + "logprob_reward": -0.013897789642214775, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.013897789642214775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0867050364613533, + 0.024818547070026398, + 0.01879589632153511, + 0.028329070657491684, + 0.031046176329255104, + 0.02072909288108349, + 0.6866887807846069, + 0.0570368729531765, + 0.02232140488922596, + 0.023529158905148506 + ], + "credit_map": [ + 0.007479914464056492, + 0.007411956321448088, + 0.014871662482619286, + 0.015354814007878304, + 0.008924007415771484, + 0.01069297268986702, + 0.010100148618221283, + 0.01186955813318491, + 0.0023008142597973347, + 0.003130906028673053, + 0.0024815620854496956, + 0.0028349985368549824, + 0.003382560098543763, + 0.00375794037245214, + 0.004049911629408598, + 0.0028798552230000496, + 0.002724192338064313, + 0.00257403077557683, + 0.002318523358553648, + 0.0026198122650384903, + 0.002489534206688404, + 0.0017021404346451163, + 0.002058973303064704, + 0.0023086892906576395, + 0.003417921019718051, + 0.0032182808499783278, + 0.00426193093881011, + 0.004827893804758787, + 0.0030988547950983047, + 0.002535640262067318, + 0.003986754454672337, + 0.0029817933682352304, + 0.003070235252380371, + 0.0018949058139696717, + 0.006013389211148024, + 0.004908462055027485, + 0.003314365865662694, + 0.0028499378822743893, + 0.004152284469455481, + 0.004842596594244242, + 0.003275336930528283, + 0.0028119951020926237, + 0.0018794245552271605, + 0.002715793903917074, + 0.002219006884843111, + 0.0030941152945160866, + 0.0026680193841457367, + 0.002065400592982769, + 0.05005985498428345, + 0.06305663287639618, + 0.019451811909675598, + 0.014189322479069233, + 0.03993495926260948, + 0.21467311680316925, + 0.0663193017244339, + 0.21900375187397003, + 0.004998022690415382, + 0.003242077073082328, + 0.016570603474974632, + 0.003306956961750984, + 0.0046524484641849995, + 0.004851491190493107, + 0.0046916441060602665, + 0.014723625965416431, + 0.004499171394854784, + 0.002657342469319701, + 0.0028217469807714224, + 0.0025454754941165447, + 0.00296065304428339, + 0.002608019858598709, + 0.002017460996285081, + 0.002211535582318902, + 0.0025831260718405247, + 0.00402793986722827, + 0.0026668596547096968, + 0.004488568287342787, + 0.0027078508865088224, + 0.002787357894703746, + 0.0016051471466198564, + 0.0026623110752552748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.013897789642214775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811623528599739 + ], + "opposite_logprob_tokens": [ + -3.288116216659546 + ], + "logprob_reward": -0.03811623528599739, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811623528599739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11502107977867126, + 0.04967738315463066, + 0.02623344212770462, + 0.033043403178453445, + 0.015815069898962975, + 0.02046489343047142, + 0.6003392338752747, + 0.04261321946978569, + 0.039734650403261185, + 0.05705762282013893 + ], + "credit_map": [ + 0.009305329993367195, + 0.010074341669678688, + 0.018268803134560585, + 0.016224462538957596, + 0.017069971188902855, + 0.011852836236357689, + 0.016086313873529434, + 0.016139009967446327, + 0.005065394099801779, + 0.005010629538446665, + 0.005805679596960545, + 0.0061501627787947655, + 0.007233862299472094, + 0.008061827160418034, + 0.0065034544095396996, + 0.0058463686145842075, + 0.002743625082075596, + 0.0027894785162061453, + 0.004559407941997051, + 0.004935586825013161, + 0.002329000039026141, + 0.0019396665738895535, + 0.00353303924202919, + 0.003403634298592806, + 0.003153779311105609, + 0.0036732275038957596, + 0.005664585158228874, + 0.005817780736833811, + 0.00316661037504673, + 0.0038721959572285414, + 0.004195922985672951, + 0.0034992981236428022, + 0.0017718671588227153, + 0.0010741855949163437, + 0.0028605153784155846, + 0.002430832479149103, + 0.0025911687407642603, + 0.0013994798064231873, + 0.001716061495244503, + 0.001970959361642599, + 0.00247020972892642, + 0.003043054835870862, + 0.0021206995006650686, + 0.002773670479655266, + 0.0019260096596553922, + 0.0028018648736178875, + 0.0023082722909748554, + 0.0030211107805371284, + 0.044524118304252625, + 0.055557478219270706, + 0.017715230584144592, + 0.01249837689101696, + 0.03467743843793869, + 0.18849778175354004, + 0.057259343564510345, + 0.18960942327976227, + 0.0049079847522079945, + 0.0042916471138596535, + 0.0093239089474082, + 0.00460978364571929, + 0.0038165615405887365, + 0.002758664544671774, + 0.005616744514554739, + 0.007287921383976936, + 0.006022972986102104, + 0.005321274045854807, + 0.004163120407611132, + 0.006963605992496014, + 0.005360433831810951, + 0.004036919213831425, + 0.0046043312177062035, + 0.0032619908452033997, + 0.007416175212711096, + 0.008656492456793785, + 0.006153385154902935, + 0.007567300461232662, + 0.00870154332369566, + 0.007309604901820421, + 0.00662342831492424, + 0.004629689734429121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.03811623528599739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205131113529205 + ], + "opposite_logprob_tokens": [ + -2.812051296234131 + ], + "logprob_reward": -0.06205131113529205, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205131113529205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10482808202505112, + 0.026883890852332115, + 0.009896868839859962, + 0.02661561779677868, + 0.011494685895740986, + 0.020653827115893364, + 0.7153913974761963, + 0.03543005883693695, + 0.018362557515501976, + 0.030443018302321434 + ], + "credit_map": [ + 0.010219849646091461, + 0.011856826953589916, + 0.015487724915146828, + 0.014594254083931446, + 0.016809947788715363, + 0.011867159977555275, + 0.011856397613883018, + 0.012135923840105534, + 0.0031626620329916477, + 0.0034646436106413603, + 0.004072093870490789, + 0.003643936477601528, + 0.0029966251458972692, + 0.0029499796219170094, + 0.003179538529366255, + 0.0034144108649343252, + 0.0012227310799062252, + 0.0012157419696450233, + 0.0018198528559878469, + 0.0014530151383951306, + 0.001049379468895495, + 0.0008802365628071129, + 0.0012662792578339577, + 0.0009896326810121536, + 0.002792075276374817, + 0.002033423399552703, + 0.004626774694770575, + 0.003730040043592453, + 0.002798465546220541, + 0.003268355969339609, + 0.004051870666444302, + 0.0033146117348223925, + 0.0014619088033214211, + 0.0009250894072465599, + 0.0017622893210500479, + 0.0014671682147309184, + 0.002168141771107912, + 0.0009729216108098626, + 0.0012127988738939166, + 0.0015243685338646173, + 0.0021495164837688208, + 0.0027524312026798725, + 0.0022261980921030045, + 0.0031071240082383156, + 0.002019523875787854, + 0.003510630689561367, + 0.0024524114560335875, + 0.0024359929375350475, + 0.0522853247821331, + 0.06592221558094025, + 0.019943075254559517, + 0.015136821195483208, + 0.04204365611076355, + 0.22377756237983704, + 0.06905658543109894, + 0.227226123213768, + 0.005134204402565956, + 0.004421516787260771, + 0.005309338215738535, + 0.003768686205148697, + 0.0036363364197313786, + 0.003766212146729231, + 0.00605013919994235, + 0.0033436277881264687, + 0.0029903592076152563, + 0.0023202463053166866, + 0.0021881426218897104, + 0.0022912933491170406, + 0.002722914796322584, + 0.0017401674995198846, + 0.001648384379222989, + 0.002461048774421215, + 0.003498064586892724, + 0.004782716277986765, + 0.0032933931797742844, + 0.003985004499554634, + 0.005314570851624012, + 0.004559549503028393, + 0.0028855265118181705, + 0.002124192425981164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.06205131113529205 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494256317615509 + ], + "opposite_logprob_tokens": [ + -2.9299426078796387 + ], + "logprob_reward": -0.05494256317615509, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494256317615509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05761668086051941, + 0.06271138042211533, + 0.020650701597332954, + 0.03393417224287987, + 0.04181501269340515, + 0.025593122467398643, + 0.4302728772163391, + 0.07188990712165833, + 0.07485855370759964, + 0.18065756559371948 + ], + "credit_map": [ + 0.006589963566511869, + 0.0065126437693834305, + 0.006068322341889143, + 0.0077847265638411045, + 0.008841370232403278, + 0.0066975802183151245, + 0.007942333817481995, + 0.007179745472967625, + 0.006623944733291864, + 0.006554134655743837, + 0.008487652987241745, + 0.00791229773312807, + 0.007186377886682749, + 0.010395807214081287, + 0.00777775002643466, + 0.00777341378852725, + 0.002734932117164135, + 0.0019771382212638855, + 0.0033813565969467163, + 0.004488825332373381, + 0.0015977052971720695, + 0.0015020504361018538, + 0.002888055983930826, + 0.002080638660117984, + 0.0036582332104444504, + 0.005403305869549513, + 0.004495228640735149, + 0.004426437430083752, + 0.003053411841392517, + 0.0032651820220053196, + 0.005826729815453291, + 0.0038056448101997375, + 0.005261414684355259, + 0.0040045129135251045, + 0.007221282925456762, + 0.005963648669421673, + 0.005950277205556631, + 0.0034645532723516226, + 0.0036934546660631895, + 0.006255868822336197, + 0.00436764769256115, + 0.00322216609492898, + 0.001968851312994957, + 0.003301584627479315, + 0.0025448647793382406, + 0.004645043984055519, + 0.0032601298298686743, + 0.0022828367073088884, + 0.030695980414748192, + 0.03977908566594124, + 0.012787139043211937, + 0.010312499478459358, + 0.025030527263879776, + 0.13278228044509888, + 0.041995707899332047, + 0.13688968122005463, + 0.00967206247150898, + 0.010035309940576553, + 0.008254744112491608, + 0.006982473656535149, + 0.007915973663330078, + 0.009417467750608921, + 0.010845636017620564, + 0.008766235783696175, + 0.023180657997727394, + 0.009372439235448837, + 0.012006189674139023, + 0.0063429828733205795, + 0.007370935752987862, + 0.006239753682166338, + 0.004444165155291557, + 0.005901436787098646, + 0.022834397852420807, + 0.039284925907850266, + 0.02343807741999626, + 0.03188120946288109, + 0.019389934837818146, + 0.02053668163716793, + 0.01240113191306591, + 0.010891230776906013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.05494256317615509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367908298969269 + ], + "opposite_logprob_tokens": [ + -3.4086790084838867 + ], + "logprob_reward": -0.03367908298969269, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367908298969269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0541439913213253, + 0.06465957313776016, + 0.013329850509762764, + 0.03801855817437172, + 0.018093863502144814, + 0.011673469096422195, + 0.7216511964797974, + 0.03220684826374054, + 0.02488313987851143, + 0.02133949287235737 + ], + "credit_map": [ + 0.004120091907680035, + 0.006259209010750055, + 0.0059473831206560135, + 0.008882759138941765, + 0.008735481649637222, + 0.008568860590457916, + 0.0048470343463122845, + 0.00678318040445447, + 0.00650701392441988, + 0.006310609169304371, + 0.00894287507981062, + 0.010293034836649895, + 0.006097814533859491, + 0.009239540435373783, + 0.008418100886046886, + 0.008850589394569397, + 0.0016431768890470266, + 0.0014840486692264676, + 0.0024815998040139675, + 0.0024669345002621412, + 0.0009065700578503311, + 0.0010680184932425618, + 0.0016886460362002254, + 0.0015908573986962438, + 0.0033738897182047367, + 0.0034692122135311365, + 0.0058219656348228455, + 0.006216037552803755, + 0.004674021620303392, + 0.004120362922549248, + 0.0058066085912287235, + 0.00453646294772625, + 0.00178648775909096, + 0.0022130468860268593, + 0.0026198290288448334, + 0.0028244226705282927, + 0.003897062735632062, + 0.0015097430441528559, + 0.0016879556933417916, + 0.0015553185949102044, + 0.0013681038981303573, + 0.001335509237833321, + 0.0010287557961419225, + 0.0019141057273373008, + 0.0011367661645635962, + 0.002294709673151374, + 0.0017087652813643217, + 0.0008867544238455594, + 0.052862003445625305, + 0.06623806804418564, + 0.019541073590517044, + 0.01602805219590664, + 0.0424061082303524, + 0.22510018944740295, + 0.06993571668863297, + 0.22954003512859344, + 0.003764071501791477, + 0.0019707472529262304, + 0.008473780006170273, + 0.0022217840887606144, + 0.002171748783439398, + 0.0016637432854622602, + 0.00366574851796031, + 0.008275224827229977, + 0.006900480017066002, + 0.0030913411173969507, + 0.0035680499859154224, + 0.0027915015816688538, + 0.0021200876217335463, + 0.0018341437680646777, + 0.001999269239604473, + 0.0025782673619687557, + 0.0030024193692952394, + 0.003968098666518927, + 0.0026279818266630173, + 0.003750303527340293, + 0.002349385991692543, + 0.0023182888980954885, + 0.0017474072519689798, + 0.0015756089705973864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.03367908298969269 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The hyperreal 3D rendering begins with a low-angle, uninterrupted, doll-like shot, gliding smoothly and silently forward across the polished, reflective white epoxy floor of a vast, sterile semiconductor manufacturing facility. A fine, clean mist wafts through the cool, filtered air, capturing and diffusing the clinical glow of the embedded LED panels overhead and the vibrant blues and ambers of status lights on the towering, complex machinery. The camera's path gracefully curves toward a multi-jointed robotic arm, its polished chrome and white composite surfaces gleaming as it extends with silent, deliberate precision. A specialized vacuum gripper at its end delicately removes a gleaming, perfectly round silicon chip from its box, its surface brimming with iridescent microscopic circuitry, refracting light into a rainbow of colors. As the robotic arm rotates and places the chip into a processing chamber, the chip's fluid journey complete. The camera pans left to reveal a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield. They stand in front of a bank of monitors with focused precision. As complex streams of data scroll across the screens, their gloved fingers methodically interact with the control interfaces, their gestures conveying absolute concentration in the facility's hushed surroundings. The camera continues its seamless forward movement, deep into the high-tech maze where human focus and automated perfection operate in a silent, synchronized ballet, creating an atmosphere of profound, futuristic tranquility and meticulous creation.", + "seed": 6344, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_7963_seed_6344_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": -0.025778696406632662, + "vlm_logprob": -0.025778696406632662, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.025778696406632662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004692138638347387 + ], + "opposite_logprob_tokens": [ + -6.004692077636719 + ], + "logprob_reward": -0.004692138638347387, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004692138638347387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07097484916448593, + 0.10273049771785736, + 0.1388399749994278, + 0.04465822875499725, + 0.04021959379315376, + 0.05802691727876663, + 0.06610871851444244, + 0.036981187760829926, + 0.10860763490200043, + 0.3328523635864258 + ], + "credit_map": [ + 0.006134535651654005, + 0.006121867801994085, + 0.012240384705364704, + 0.009204152971506119, + 0.007556614466011524, + 0.007233093958348036, + 0.012562376447021961, + 0.009921824559569359, + 0.010094362311065197, + 0.009722133167088032, + 0.010687416419386864, + 0.014058982022106647, + 0.012716943398118019, + 0.015969550237059593, + 0.01783588156104088, + 0.011645225808024406, + 0.006955935154110193, + 0.005224319640547037, + 0.006892832927405834, + 0.0059274123050272465, + 0.006925826892256737, + 0.007372790947556496, + 0.049047715961933136, + 0.05049314722418785, + 0.002910058246925473, + 0.003605414181947708, + 0.0035644888412207365, + 0.006485659163445234, + 0.008649999275803566, + 0.006655325181782246, + 0.006715917959809303, + 0.006071365438401699, + 0.008316995576024055, + 0.006076495163142681, + 0.0048341406509280205, + 0.004068461712449789, + 0.0039469292387366295, + 0.006460653617978096, + 0.00289165903814137, + 0.003624259028583765, + 0.009295188821852207, + 0.005592950154095888, + 0.016989851370453835, + 0.004174649715423584, + 0.0038819245528429747, + 0.005070559214800596, + 0.0031959954649209976, + 0.00982579868286848, + 0.005991902202367783, + 0.0062020584009587765, + 0.01068069040775299, + 0.006890334188938141, + 0.007909540086984634, + 0.012001559138298035, + 0.009893815033137798, + 0.006538813933730125, + 0.004608477931469679, + 0.007122375536710024, + 0.0038752162363380194, + 0.00541574414819479, + 0.006415002513676882, + 0.002538690809160471, + 0.003432154655456543, + 0.003573525696992874, + 0.010200279764831066, + 0.022884057834744453, + 0.01545944157987833, + 0.013953392393887043, + 0.015745559707283974, + 0.01217213086783886, + 0.010288841091096401, + 0.0079039316624403, + 0.06274410337209702, + 0.049879662692546844, + 0.09350082278251648, + 0.04722162336111069, + 0.015227993950247765, + 0.017749391496181488, + 0.032667502760887146, + 0.013861270621418953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.004692138638347387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255324762314558 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.01255324762314558, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255324762314558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09689868241548538, + 0.10699222981929779, + 0.2810918986797333, + 0.060687076300382614, + 0.03738737106323242, + 0.06183620169758797, + 0.04949219897389412, + 0.04631655290722847, + 0.10786289721727371, + 0.15143483877182007 + ], + "credit_map": [ + 0.008642171509563923, + 0.010581567883491516, + 0.016432711854577065, + 0.013335717841982841, + 0.015120978467166424, + 0.008833151310682297, + 0.013818350620567799, + 0.010134036652743816, + 0.015270419418811798, + 0.01068855356425047, + 0.011115101166069508, + 0.013989324681460857, + 0.01310854870826006, + 0.011825643479824066, + 0.013885577209293842, + 0.01710907183587551, + 0.0050024292431771755, + 0.002792751183733344, + 0.004469647537916899, + 0.008015927858650684, + 0.007207944057881832, + 0.006994821596890688, + 0.12311767041683197, + 0.12349074333906174, + 0.008417036384344101, + 0.00696147233247757, + 0.004759257193654776, + 0.011661631055176258, + 0.008924775756895542, + 0.006718097720295191, + 0.006355307064950466, + 0.006889502052217722, + 0.004790333099663258, + 0.0037468711379915476, + 0.003486210945993662, + 0.007791677489876747, + 0.00667064031586051, + 0.004252093844115734, + 0.0032171206548810005, + 0.003432427765801549, + 0.008361153304576874, + 0.007198437117040157, + 0.020496375858783722, + 0.00453274417668581, + 0.003885463811457157, + 0.005237291567027569, + 0.003716319566592574, + 0.008408418856561184, + 0.006419767159968615, + 0.0031141946092247963, + 0.00650322949513793, + 0.0037189519498497248, + 0.005267590284347534, + 0.006312001962214708, + 0.013464031741023064, + 0.004692435264587402, + 0.006299030035734177, + 0.009003703482449055, + 0.00432599289342761, + 0.006261372473090887, + 0.008832133375108242, + 0.0035990672186017036, + 0.0037025748752057552, + 0.004292682278901339, + 0.013764786534011364, + 0.014841361902654171, + 0.020833170041441917, + 0.017358431592583656, + 0.0111231105402112, + 0.011278891935944557, + 0.010157663375139236, + 0.008505486883223057, + 0.018169566988945007, + 0.0287737175822258, + 0.02702413871884346, + 0.0272990632802248, + 0.012333352118730545, + 0.012225993908941746, + 0.016936803236603737, + 0.00867222249507904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.01255324762314558 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290520533919334 + ], + "opposite_logprob_tokens": [ + -3.7732906341552734 + ], + "logprob_reward": -0.023290520533919334, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290520533919334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.167927086353302, + 0.17000190913677216, + 0.08311473578214645, + 0.037974148988723755, + 0.07886417955160141, + 0.0858614593744278, + 0.07539937645196915, + 0.11641594022512436, + 0.03845541179180145, + 0.1459856778383255 + ], + "credit_map": [ + 0.011481710709631443, + 0.01595216430723667, + 0.020942222326993942, + 0.017392262816429138, + 0.016150634735822678, + 0.0086503392085433, + 0.043647486716508865, + 0.0337102897465229, + 0.01650822162628174, + 0.018122417852282524, + 0.02019716240465641, + 0.026179084554314613, + 0.02372181974351406, + 0.026597704738378525, + 0.020505817607045174, + 0.018169699236750603, + 0.007324223406612873, + 0.007797989994287491, + 0.005628120619803667, + 0.006298778112977743, + 0.00619980413466692, + 0.007992207072675228, + 0.020471666008234024, + 0.021401960402727127, + 0.004782671574503183, + 0.004333422984927893, + 0.0032803898211568594, + 0.0060618226416409016, + 0.004066201858222485, + 0.0051106917671859264, + 0.0049554165452718735, + 0.005383533425629139, + 0.011088022030889988, + 0.011429831385612488, + 0.00658205384388566, + 0.022120967507362366, + 0.006091221701353788, + 0.009605764411389828, + 0.0071081374771893024, + 0.004838190041482449, + 0.01278256718069315, + 0.00936883594840765, + 0.011411221697926521, + 0.006725855637341738, + 0.009280589409172535, + 0.015535041689872742, + 0.0075253695249557495, + 0.01323198527097702, + 0.006033178418874741, + 0.005341807845979929, + 0.016179127618670464, + 0.0068235271610319614, + 0.00611145468428731, + 0.0076079717837274075, + 0.02063308283686638, + 0.006669237278401852, + 0.012109294533729553, + 0.024219335988163948, + 0.011228435672819614, + 0.019013555720448494, + 0.02311249077320099, + 0.008770810440182686, + 0.008507844991981983, + 0.009454180486500263, + 0.0037568763364106417, + 0.006743879057466984, + 0.005576211027801037, + 0.003937741741538048, + 0.007090444210916758, + 0.004309356212615967, + 0.0037032777909189463, + 0.0033376284409314394, + 0.021501652896404266, + 0.022531935945153236, + 0.023106159642338753, + 0.02315153367817402, + 0.019147349521517754, + 0.011659754440188408, + 0.01493001263588667, + 0.00995729025453329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.023290520533919334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062028564512729645 + ], + "opposite_logprob_tokens": [ + -2.812028646469116 + ], + "logprob_reward": -0.062028564512729645, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062028564512729645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022134920582175255, + 0.027849644422531128, + 0.057748034596443176, + 0.015501754358410835, + 0.033180978149175644, + 0.029165398329496384, + 0.023244181647896767, + 0.07412208616733551, + 0.10064635425806046, + 0.6164066195487976 + ], + "credit_map": [ + 0.0021645501255989075, + 0.002257764572277665, + 0.0030929625499993563, + 0.0027995745185762644, + 0.002055918565019965, + 0.0019800616428256035, + 0.003942747134715319, + 0.0038413412403315306, + 0.0027090159710496664, + 0.0031905153300613165, + 0.004572747275233269, + 0.005622304975986481, + 0.0030470085330307484, + 0.0032641866710036993, + 0.0032735487911850214, + 0.002170316642150283, + 0.002041194587945938, + 0.0020895646885037422, + 0.0018790693720802665, + 0.0025975543539971113, + 0.00314727914519608, + 0.0030521899461746216, + 0.021432682871818542, + 0.02150850184261799, + 0.0018035753164440393, + 0.0016000490868464112, + 0.0016029555117711425, + 0.0029802185017615557, + 0.0016220291145145893, + 0.0015340748941525817, + 0.0021245761308819056, + 0.002234275918453932, + 0.004680065903812647, + 0.002749019768089056, + 0.0028078898321837187, + 0.003974416293203831, + 0.007592609152197838, + 0.0042344857938587666, + 0.0025763690937310457, + 0.004566122777760029, + 0.00574622955173254, + 0.002870906377211213, + 0.008756151422858238, + 0.002017026534304023, + 0.002272587036713958, + 0.0024539190344512463, + 0.0017324108630418777, + 0.003316169837489724, + 0.002058799611404538, + 0.0023167915642261505, + 0.0031825522892177105, + 0.0023854675237089396, + 0.003375428728759289, + 0.005387603770941496, + 0.0020358182955533266, + 0.002501719631254673, + 0.007629093714058399, + 0.01426314190030098, + 0.008165172301232815, + 0.01232035644352436, + 0.013014817610383034, + 0.005512291565537453, + 0.006548695266246796, + 0.0066685243509709835, + 0.01132710836827755, + 0.018402282148599625, + 0.016733886674046516, + 0.013693488202989101, + 0.013521797955036163, + 0.009211668744683266, + 0.009508561342954636, + 0.008247559890151024, + 0.13894884288311005, + 0.07954277098178864, + 0.17214444279670715, + 0.07783552259206772, + 0.031822893768548965, + 0.031939100474119186, + 0.06190655380487442, + 0.022266512736678123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.062028564512729645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026329010725021362 + ], + "opposite_logprob_tokens": [ + -3.6513290405273438 + ], + "logprob_reward": -0.026329010725021362, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026329010725021362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14171326160430908, + 0.11923278868198395, + 0.24994264543056488, + 0.043422214686870575, + 0.033094801008701324, + 0.06189264729619026, + 0.05452742427587509, + 0.023023143410682678, + 0.18532001972198486, + 0.087831050157547 + ], + "credit_map": [ + 0.012507300823926926, + 0.01061831321567297, + 0.01724916137754917, + 0.017221879214048386, + 0.012625419534742832, + 0.0093718646094203, + 0.03621004521846771, + 0.025909291580319405, + 0.00938135851174593, + 0.010602694004774094, + 0.01462967786937952, + 0.018510567024350166, + 0.009878088720142841, + 0.010501719079911709, + 0.024720344692468643, + 0.021008333191275597, + 0.008589635603129864, + 0.0096356812864542, + 0.012510930188000202, + 0.012911573052406311, + 0.011504238471388817, + 0.010913950391113758, + 0.09197662025690079, + 0.09190001338720322, + 0.003929029684513807, + 0.004692987073212862, + 0.003693660721182823, + 0.005105136428028345, + 0.0070331403985619545, + 0.0069512478075921535, + 0.007056864909827709, + 0.004960147198289633, + 0.003740157699212432, + 0.003208054229617119, + 0.0031991133000701666, + 0.004869516473263502, + 0.005046699196100235, + 0.006391489412635565, + 0.0028820736333727837, + 0.0037576970644295216, + 0.011338161304593086, + 0.008924318477511406, + 0.01511242426931858, + 0.004674897529184818, + 0.005060593131929636, + 0.009452778846025467, + 0.004265069030225277, + 0.0030644037760794163, + 0.004511683247983456, + 0.004033713601529598, + 0.009415150620043278, + 0.00483949389308691, + 0.0058403112925589085, + 0.007442764472216368, + 0.013457149267196655, + 0.004987157415598631, + 0.0029875009786337614, + 0.0028838254511356354, + 0.0029740589670836926, + 0.003106784773990512, + 0.00321531156077981, + 0.0032244364265352488, + 0.002137948991730809, + 0.002493276260793209, + 0.012780444696545601, + 0.04913176968693733, + 0.020424572750926018, + 0.017970038577914238, + 0.0434153713285923, + 0.018560107797384262, + 0.012897295877337456, + 0.010140424594283104, + 0.01128324680030346, + 0.016288872808218002, + 0.01959158107638359, + 0.007844456471502781, + 0.009907543659210205, + 0.008080530911684036, + 0.009737402200698853, + 0.005097412038594484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.026329010725021362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The hyperreal 3D rendering begins with a low-angle, uninterrupted, doll-like shot, gliding smoothly and silently forward across the polished, reflective white epoxy floor of a vast, sterile semiconductor manufacturing facility. A fine, clean mist wafts through the cool, filtered air, capturing and diffusing the clinical glow of the embedded LED panels overhead and the vibrant blues and ambers of status lights on the towering, complex machinery. The camera's path gracefully curves toward a multi-jointed robotic arm, its polished chrome and white composite surfaces gleaming as it extends with silent, deliberate precision. A specialized vacuum gripper at its end delicately removes a gleaming, perfectly round silicon chip from its box, its surface brimming with iridescent microscopic circuitry, refracting light into a rainbow of colors. As the robotic arm rotates and places the chip into a processing chamber, the chip's fluid journey complete. The camera pans left to reveal a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield. They stand in front of a bank of monitors with focused precision. As complex streams of data scroll across the screens, their gloved fingers methodically interact with the control interfaces, their gestures conveying absolute concentration in the facility's hushed surroundings. The camera continues its seamless forward movement, deep into the high-tech maze where human focus and automated perfection operate in a silent, synchronized ballet, creating an atmosphere of profound, futuristic tranquility and meticulous creation.", + "seed": 6345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_7963_seed_6345_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": -0.07530314773321152, + "vlm_logprob": -0.07530314773321152, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.07530314773321152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009565472602844238 + ], + "opposite_logprob_tokens": [ + -5.009565353393555 + ], + "logprob_reward": -0.009565472602844238, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009565472602844238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2016189694404602, + 0.10535749793052673, + 0.10852349549531937, + 0.19351725280284882, + 0.10351469367742538, + 0.07447315752506256, + 0.05717720463871956, + 0.049878280609846115, + 0.03641702979803085, + 0.06952238082885742 + ], + "credit_map": [ + 0.02572662942111492, + 0.020477578043937683, + 0.028276760131120682, + 0.030144991353154182, + 0.020304054021835327, + 0.01959379017353058, + 0.02404184080660343, + 0.03305332735180855, + 0.008797467686235905, + 0.00953739881515503, + 0.011483607813715935, + 0.012524699792265892, + 0.008911298587918282, + 0.009551393799483776, + 0.020212572067975998, + 0.024339064955711365, + 0.013555670157074928, + 0.01413158979266882, + 0.014134109951555729, + 0.01382648665457964, + 0.014013862237334251, + 0.010563709773123264, + 0.014094959944486618, + 0.014203105121850967, + 0.017067089676856995, + 0.015505707822740078, + 0.051849525421857834, + 0.013640441000461578, + 0.01946817897260189, + 0.022120961919426918, + 0.04500989615917206, + 0.008855456486344337, + 0.0098800640553236, + 0.01149611733853817, + 0.011516704224050045, + 0.009694362990558147, + 0.021679870784282684, + 0.01584652066230774, + 0.012606683187186718, + 0.01079437043517828, + 0.010434720665216446, + 0.006738733965903521, + 0.011957700364291668, + 0.011892778798937798, + 0.006075324956327677, + 0.008194361813366413, + 0.01182233914732933, + 0.007357198745012283, + 0.004691144917160273, + 0.00594366854056716, + 0.008459423668682575, + 0.007669923361390829, + 0.008127976208925247, + 0.007367503363639116, + 0.007847939617931843, + 0.007069624029099941, + 0.00569147290661931, + 0.005280630197376013, + 0.005235913675278425, + 0.009501367807388306, + 0.005377122666686773, + 0.004908546339720488, + 0.006520337425172329, + 0.007362889591604471, + 0.005198929458856583, + 0.00417284993454814, + 0.003896478796377778, + 0.0061301435343921185, + 0.003355707973241806, + 0.004345024935901165, + 0.005648821592330933, + 0.0036690752021968365, + 0.009186607785522938, + 0.010958989150822163, + 0.009854684583842754, + 0.010870922356843948, + 0.0049073342233896255, + 0.005470814649015665, + 0.012694811448454857, + 0.005578218027949333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.009565472602844238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371262177824974 + ], + "opposite_logprob_tokens": [ + -3.40871262550354 + ], + "logprob_reward": -0.03371262177824974, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371262177824974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1234404593706131, + 0.1475546509027481, + 0.1393338292837143, + 0.31231266260147095, + 0.08210191875696182, + 0.01929726079106331, + 0.025278601795434952, + 0.07076190412044525, + 0.016759280115365982, + 0.06315942853689194 + ], + "credit_map": [ + 0.018651016056537628, + 0.014379322528839111, + 0.009888053871691227, + 0.01285264641046524, + 0.01727992482483387, + 0.013892745599150658, + 0.020059863105416298, + 0.01643688790500164, + 0.016873503103852272, + 0.027720540761947632, + 0.011457656510174274, + 0.016741899773478508, + 0.020775798708200455, + 0.015644805505871773, + 0.017897840589284897, + 0.020442599430680275, + 0.022894704714417458, + 0.018496476113796234, + 0.017437804490327835, + 0.017023250460624695, + 0.015363515354692936, + 0.014622525312006474, + 0.017036912962794304, + 0.016458643600344658, + 0.01642351597547531, + 0.017799967899918556, + 0.02429349534213543, + 0.024118756875395775, + 0.031924404203891754, + 0.07254450768232346, + 0.10901038348674774, + 0.016197623685002327, + 0.006544430740177631, + 0.0069668227806687355, + 0.013525507412850857, + 0.013723151758313179, + 0.010367915034294128, + 0.009279975667595863, + 0.013027982786297798, + 0.008666140027344227, + 0.0023410762660205364, + 0.0014749598922207952, + 0.001650008955039084, + 0.00286895246244967, + 0.0025803903117775917, + 0.0031287490855902433, + 0.0031293013598769903, + 0.0021238219924271107, + 0.0018596156733110547, + 0.0023857110645622015, + 0.003032959531992674, + 0.0029948409646749496, + 0.003438840387389064, + 0.0053017581813037395, + 0.0026462057139724493, + 0.003618670394644141, + 0.0072669596411287785, + 0.00609154487028718, + 0.015447951853275299, + 0.02370573580265045, + 0.002589393174275756, + 0.0054293530993163586, + 0.005593778565526009, + 0.0046371822245419025, + 0.002448580227792263, + 0.0023781585041433573, + 0.001974237384274602, + 0.0026164811570197344, + 0.002081410028040409, + 0.001560322241857648, + 0.0022868039086461067, + 0.0014132873620837927, + 0.01068483479321003, + 0.007079120725393295, + 0.011651908978819847, + 0.005644537042826414, + 0.005981619004160166, + 0.006201837211847305, + 0.009472528472542763, + 0.00644303485751152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.03371262177824974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13876855373382568, + 0.08445686101913452, + 0.1166314035654068, + 0.16006553173065186, + 0.10622387379407883, + 0.054707854986190796, + 0.06580697745084763, + 0.13405075669288635, + 0.0454997792840004, + 0.09378840774297714 + ], + "credit_map": [ + 0.013166632503271103, + 0.017672758549451828, + 0.0208357572555542, + 0.020944364368915558, + 0.013741561211645603, + 0.012422047555446625, + 0.023106269538402557, + 0.01687915436923504, + 0.00825580209493637, + 0.009130784310400486, + 0.006629178766161203, + 0.011322684586048126, + 0.006501508876681328, + 0.0075837476179003716, + 0.016861876472830772, + 0.01817127875983715, + 0.01496727205812931, + 0.017688076943159103, + 0.010461932979524136, + 0.016435813158750534, + 0.01342481467872858, + 0.01174590177834034, + 0.015836307778954506, + 0.01607128232717514, + 0.016728127375245094, + 0.01890389248728752, + 0.034526340663433075, + 0.01414320059120655, + 0.01797858253121376, + 0.016861865296959877, + 0.03070095181465149, + 0.010222576558589935, + 0.007561418693512678, + 0.012568101286888123, + 0.010250438936054707, + 0.008589296601712704, + 0.021408598870038986, + 0.01662389561533928, + 0.01821129396557808, + 0.011010833084583282, + 0.006735304836183786, + 0.003884984413161874, + 0.0048885284923017025, + 0.007526074536144733, + 0.006543109659105539, + 0.008946680463850498, + 0.00970936007797718, + 0.006473815068602562, + 0.005116583313792944, + 0.00664119515568018, + 0.007284902967512608, + 0.008885612711310387, + 0.00779667217284441, + 0.012554729357361794, + 0.0092503372579813, + 0.008276942186057568, + 0.014817686751484871, + 0.011811912059783936, + 0.032286420464515686, + 0.042767759412527084, + 0.006538162939250469, + 0.009106580168008804, + 0.009192392230033875, + 0.007529845926910639, + 0.0072474111802875996, + 0.004403585102409124, + 0.0049621835350990295, + 0.005362808238714933, + 0.005046884994953871, + 0.004965158645063639, + 0.009001132100820541, + 0.0045106131583452225, + 0.015699708834290504, + 0.010667391121387482, + 0.017023790627717972, + 0.011142797768115997, + 0.009701957926154137, + 0.00634555472061038, + 0.01566721498966217, + 0.007539987098425627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.04306614026427269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252091646194458 + ], + "opposite_logprob_tokens": [ + -1.502091646194458 + ], + "logprob_reward": -0.252091646194458, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252091646194458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16300667822360992, + 0.08532138913869858, + 0.1548999398946762, + 0.16706065833568573, + 0.19387245178222656, + 0.052855104207992554, + 0.03611360862851143, + 0.06070402264595032, + 0.024450652301311493, + 0.061715517193078995 + ], + "credit_map": [ + 0.022483687847852707, + 0.016796959564089775, + 0.0220064464956522, + 0.021773049607872963, + 0.015671296045184135, + 0.01701235957443714, + 0.01857548952102661, + 0.028687387704849243, + 0.01047635730355978, + 0.011709067970514297, + 0.006955766584724188, + 0.009217677637934685, + 0.008017807267606258, + 0.007841482758522034, + 0.015617971308529377, + 0.01548525970429182, + 0.029197219759225845, + 0.01974841021001339, + 0.017828311771154404, + 0.015418661758303642, + 0.01814839243888855, + 0.01700741797685623, + 0.019295303151011467, + 0.018256226554512978, + 0.009480824694037437, + 0.013744192197918892, + 0.02667432650923729, + 0.0133219538256526, + 0.019435951486229897, + 0.02192840538918972, + 0.05131719633936882, + 0.011157813481986523, + 0.01568700559437275, + 0.022922787815332413, + 0.021332522854208946, + 0.0220290906727314, + 0.04493800550699234, + 0.023701520636677742, + 0.02545269951224327, + 0.017808813601732254, + 0.008389389142394066, + 0.0051324074156582355, + 0.005694021936506033, + 0.007069801911711693, + 0.006169910077005625, + 0.006231036968529224, + 0.008900459855794907, + 0.005268072709441185, + 0.002912225667387247, + 0.0035630613565444946, + 0.0044311340898275375, + 0.0043843393214046955, + 0.0047990428283810616, + 0.007323979865759611, + 0.003898038063198328, + 0.004801788832992315, + 0.006004075985401869, + 0.006496991030871868, + 0.006960817147046328, + 0.012751249596476555, + 0.005867870524525642, + 0.006553273648023605, + 0.007921600714325905, + 0.00814814306795597, + 0.0034508174285292625, + 0.002049437491223216, + 0.003668254241347313, + 0.0026416999753564596, + 0.0025153723545372486, + 0.0024985852651298046, + 0.003976303618401289, + 0.0036501805298030376, + 0.010163982398808002, + 0.006638703867793083, + 0.01345226913690567, + 0.009313282556831837, + 0.0062896027229726315, + 0.0036652551498264074, + 0.007860136218369007, + 0.004332279786467552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.252091646194458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807985782623291 + ], + "opposite_logprob_tokens": [ + -3.2880797386169434 + ], + "logprob_reward": -0.03807985782623291, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807985782623291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1365249902009964, + 0.2086736410856247, + 0.08394458144903183, + 0.08750902116298676, + 0.06769157201051712, + 0.2115841954946518, + 0.06535577028989792, + 0.0500536784529686, + 0.05044639855623245, + 0.038216181099414825 + ], + "credit_map": [ + 0.01735401712357998, + 0.024628708139061928, + 0.015835188329219818, + 0.013070176355540752, + 0.01832110993564129, + 0.0161904264241457, + 0.011695741675794125, + 0.01942961849272251, + 0.02381943352520466, + 0.030646126717329025, + 0.012659097090363503, + 0.02055077813565731, + 0.02294742874801159, + 0.018654420971870422, + 0.038704000413417816, + 0.040692344307899475, + 0.016443436965346336, + 0.011522437445819378, + 0.010746379382908344, + 0.007367096841335297, + 0.00893771555274725, + 0.009137757122516632, + 0.00911161582916975, + 0.010678143240511417, + 0.006460174918174744, + 0.006526286713778973, + 0.02360486052930355, + 0.005775401834398508, + 0.00866225641220808, + 0.009486449882388115, + 0.02045449987053871, + 0.0065390910021960735, + 0.004462873097509146, + 0.009137698449194431, + 0.0083268778398633, + 0.0103067671880126, + 0.01171114668250084, + 0.005672698840498924, + 0.0071757473051548, + 0.010897761210799217, + 0.03051234409213066, + 0.017860811203718185, + 0.038910940289497375, + 0.032483961433172226, + 0.014249291270971298, + 0.01776948943734169, + 0.03916230425238609, + 0.02063503861427307, + 0.006737561896443367, + 0.005880066193640232, + 0.0104093998670578, + 0.00740841357037425, + 0.008571941405534744, + 0.011152208782732487, + 0.006818265188485384, + 0.008377916179597378, + 0.005959715228527784, + 0.005866508465260267, + 0.005716858897358179, + 0.008971991017460823, + 0.005122708156704903, + 0.0046593728475272655, + 0.008749434724450111, + 0.00500708632171154, + 0.005369509104639292, + 0.004004949703812599, + 0.005260390229523182, + 0.006805070210248232, + 0.006326506845653057, + 0.009786829352378845, + 0.006132819689810276, + 0.006760322954505682, + 0.005351252853870392, + 0.004842050839215517, + 0.007683201227337122, + 0.006301978137344122, + 0.0033337315544486046, + 0.0023016228806227446, + 0.004979383200407028, + 0.0034229604061692953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.03807985782623291 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The hyperreal 3D rendering begins with a low-angle, uninterrupted, doll-like shot, gliding smoothly and silently forward across the polished, reflective white epoxy floor of a vast, sterile semiconductor manufacturing facility. A fine, clean mist wafts through the cool, filtered air, capturing and diffusing the clinical glow of the embedded LED panels overhead and the vibrant blues and ambers of status lights on the towering, complex machinery. The camera's path gracefully curves toward a multi-jointed robotic arm, its polished chrome and white composite surfaces gleaming as it extends with silent, deliberate precision. A specialized vacuum gripper at its end delicately removes a gleaming, perfectly round silicon chip from its box, its surface brimming with iridescent microscopic circuitry, refracting light into a rainbow of colors. As the robotic arm rotates and places the chip into a processing chamber, the chip's fluid journey complete. The camera pans left to reveal a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield. They stand in front of a bank of monitors with focused precision. As complex streams of data scroll across the screens, their gloved fingers methodically interact with the control interfaces, their gestures conveying absolute concentration in the facility's hushed surroundings. The camera continues its seamless forward movement, deep into the high-tech maze where human focus and automated perfection operate in a silent, synchronized ballet, creating an atmosphere of profound, futuristic tranquility and meticulous creation.", + "seed": 6346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_3/index_7963_seed_6346_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": -0.08996562622487544, + "vlm_logprob": -0.08996562622487544, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.08996562622487544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0178073700517416 + ], + "opposite_logprob_tokens": [ + -4.517807483673096 + ], + "logprob_reward": -0.0178073700517416, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0178073700517416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06792207062244415, + 0.036874689161777496, + 0.1246730238199234, + 0.050442226231098175, + 0.017119556665420532, + 0.03612793982028961, + 0.044433631002902985, + 0.04092477262020111, + 0.5556142330169678, + 0.025867873802781105 + ], + "credit_map": [ + 0.00807274878025055, + 0.006542319431900978, + 0.012690541334450245, + 0.007903163321316242, + 0.005435032304376364, + 0.008744856342673302, + 0.008212000131607056, + 0.010321409441530704, + 0.0031801750883460045, + 0.003521726932376623, + 0.00700361980125308, + 0.0062631890177726746, + 0.0030217862222343683, + 0.004270466044545174, + 0.0047284518368542194, + 0.004885274451225996, + 0.014039376750588417, + 0.01720324717462063, + 0.018522080034017563, + 0.012727631255984306, + 0.015286725014448166, + 0.01846824586391449, + 0.013143653981387615, + 0.01528205443173647, + 0.006176074501127005, + 0.006441671401262283, + 0.008343664929270744, + 0.005663846619427204, + 0.005016825627535582, + 0.006872081197798252, + 0.0056235650554299355, + 0.006304497830569744, + 0.00194950879085809, + 0.001965726027265191, + 0.002099469769746065, + 0.001970849232748151, + 0.002012044657021761, + 0.0016605979762971401, + 0.0020683826878666878, + 0.0033929769415408373, + 0.004076652228832245, + 0.0032814969308674335, + 0.00320670404471457, + 0.004485337994992733, + 0.0031179096549749374, + 0.002886479953303933, + 0.009563514031469822, + 0.005509845446795225, + 0.00588027760386467, + 0.00424985634163022, + 0.0027622119523584843, + 0.0051742009818553925, + 0.0076423464342951775, + 0.004946967586874962, + 0.010846939869225025, + 0.002930826973170042, + 0.005423582158982754, + 0.005353661719709635, + 0.005383494775742292, + 0.004233984276652336, + 0.004791969899088144, + 0.00471990741789341, + 0.005525666289031506, + 0.005492508411407471, + 0.016967736184597015, + 0.21037878096103668, + 0.012212352827191353, + 0.21783290803432465, + 0.02214086800813675, + 0.01399336475878954, + 0.03656354546546936, + 0.025524640455842018, + 0.0029975478537380695, + 0.0023453652393072844, + 0.002974833594635129, + 0.004096673801541328, + 0.003272400237619877, + 0.003331955289468169, + 0.0029206855688244104, + 0.003928412217646837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.0178073700517416 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371642529964447 + ], + "opposite_logprob_tokens": [ + -3.4087164402008057 + ], + "logprob_reward": -0.03371642529964447, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371642529964447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15115512907505035, + 0.04802927002310753, + 0.1047806441783905, + 0.024190753698349, + 0.016734713688492775, + 0.027399182319641113, + 0.03046039119362831, + 0.022452758625149727, + 0.5352843403816223, + 0.03951287269592285 + ], + "credit_map": [ + 0.018477309495210648, + 0.014656273648142815, + 0.02483823522925377, + 0.014465655200183392, + 0.018453244119882584, + 0.018737899139523506, + 0.01866469532251358, + 0.022861827164888382, + 0.005007627420127392, + 0.00432857358828187, + 0.008325846865773201, + 0.007389827631413937, + 0.004524307791143656, + 0.005542962346225977, + 0.006704376079142094, + 0.006205747369676828, + 0.008739807642996311, + 0.013330036774277687, + 0.015992607921361923, + 0.011571021750569344, + 0.021306289359927177, + 0.013859013095498085, + 0.009985785000026226, + 0.00999608263373375, + 0.002408790634945035, + 0.0029762624762952328, + 0.003505876986309886, + 0.0033976451959460974, + 0.002360407030209899, + 0.002401150530204177, + 0.004121212288737297, + 0.003019405994564295, + 0.0017660657176747918, + 0.001686847535893321, + 0.0031027693767100573, + 0.0020246149506419897, + 0.0015088168438524008, + 0.0020684534683823586, + 0.001861270167864859, + 0.0027158751618117094, + 0.0026307322550565004, + 0.0020062108524143696, + 0.002245442010462284, + 0.0036553428508341312, + 0.00293390778824687, + 0.003826608182862401, + 0.006308598909527063, + 0.003792340401560068, + 0.00479670986533165, + 0.003704304341226816, + 0.0018451509531587362, + 0.003256196388974786, + 0.0047886245884001255, + 0.003358449088409543, + 0.006312017794698477, + 0.0023989384062588215, + 0.003048145677894354, + 0.003037855727598071, + 0.0034519026521593332, + 0.002428603358566761, + 0.002261657966300845, + 0.00273302779532969, + 0.002723495941609144, + 0.0027680692728608847, + 0.01717870868742466, + 0.2006983906030655, + 0.012960430234670639, + 0.2110990583896637, + 0.020175164565443993, + 0.01272154226899147, + 0.03571045398712158, + 0.02474062144756317, + 0.0038645765744149685, + 0.005781378131359816, + 0.004243227653205395, + 0.006693992298096418, + 0.007572122849524021, + 0.003656736807897687, + 0.004302359651774168, + 0.003398479660972953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.03371642529964447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045708298683167 + ], + "opposite_logprob_tokens": [ + -2.8120458126068115 + ], + "logprob_reward": -0.062045708298683167, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045708298683167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08781024068593979, + 0.04617200791835785, + 0.12141038477420807, + 0.04880747199058533, + 0.026884004473686218, + 0.022755248472094536, + 0.023109450936317444, + 0.03442547470331192, + 0.5468524694442749, + 0.04177318513393402 + ], + "credit_map": [ + 0.011124796234071255, + 0.008320613764226437, + 0.011327296495437622, + 0.01405850239098072, + 0.0069235144183039665, + 0.011154278181493282, + 0.011996139772236347, + 0.012905108742415905, + 0.004689161200076342, + 0.0046781981363892555, + 0.0064904456958174706, + 0.007441639434546232, + 0.0038645872846245766, + 0.006765856873244047, + 0.006002781447023153, + 0.006239340174943209, + 0.012273234315216541, + 0.016311850398778915, + 0.016365651041269302, + 0.012725595384836197, + 0.017834877595305443, + 0.019012026488780975, + 0.013262300752103329, + 0.013624873012304306, + 0.006847627926617861, + 0.005355847533792257, + 0.007259889505803585, + 0.005307994317263365, + 0.004071890842169523, + 0.006701333913952112, + 0.005886167287826538, + 0.007376726251095533, + 0.00216777715831995, + 0.002636364893987775, + 0.0025600872468203306, + 0.0037598533090204, + 0.002184672048315406, + 0.0022679453250020742, + 0.004061209969222546, + 0.0072460961528122425, + 0.0021634111180901527, + 0.001834013732150197, + 0.0020220514852553606, + 0.0024604899808764458, + 0.002282129367813468, + 0.0034475205466151237, + 0.005391835235059261, + 0.0031537977047264576, + 0.002533555030822754, + 0.0029107248410582542, + 0.0016220171237364411, + 0.0031318049877882004, + 0.003119903616607189, + 0.0026173267979174852, + 0.00516326492652297, + 0.0020108562894165516, + 0.004752140026539564, + 0.004881365690380335, + 0.004913997370749712, + 0.0032114116474986076, + 0.004157673101872206, + 0.004180428106337786, + 0.0032355457078665495, + 0.005092917010188103, + 0.01730303466320038, + 0.20316648483276367, + 0.013696392998099327, + 0.21863782405853271, + 0.019971009343862534, + 0.012797796167433262, + 0.03654342144727707, + 0.024736566469073296, + 0.005341339390724897, + 0.004416143987327814, + 0.004917659796774387, + 0.007492215372622013, + 0.006499483250081539, + 0.0048451293259859085, + 0.003910669591277838, + 0.004350551404058933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.062045708298683167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813270092010498 + ], + "opposite_logprob_tokens": [ + -1.4063270092010498 + ], + "logprob_reward": -0.2813270092010498, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813270092010498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045907001942396164, + 0.022891096770763397, + 0.018726620823144913, + 0.006908646319061518, + 0.008519114926457405, + 0.015344512648880482, + 0.03271247446537018, + 0.015555953606963158, + 0.8245707154273987, + 0.008863831870257854 + ], + "credit_map": [ + 0.006309473887085915, + 0.004598760511726141, + 0.00769012188538909, + 0.005424057599157095, + 0.003740194719284773, + 0.005069251172244549, + 0.00563597260043025, + 0.007439176086336374, + 0.0021037200931459665, + 0.002239328809082508, + 0.003134215949103236, + 0.0025157700292766094, + 0.002845852868631482, + 0.003806274849921465, + 0.002833904232829809, + 0.0034120329655706882, + 0.002592582255601883, + 0.0025964214000850916, + 0.002318935003131628, + 0.0021825230214744806, + 0.001984271453693509, + 0.002543061040341854, + 0.002297238912433386, + 0.002211590763181448, + 0.0005903841811232269, + 0.0007233608630485833, + 0.0008676082361489534, + 0.0010031871497631073, + 0.0006864245515316725, + 0.0006322978879325092, + 0.0012404643930494785, + 0.0011649199295789003, + 0.0009439830901101232, + 0.000959686585702002, + 0.0008991591748781502, + 0.0011132860090583563, + 0.0009580309269949794, + 0.0009296082425862551, + 0.0010473082074895501, + 0.0016680534463375807, + 0.0014871209859848022, + 0.001356899505481124, + 0.0013574676122516394, + 0.0013251922791823745, + 0.0016003097407519817, + 0.0023977395612746477, + 0.0035887574777007103, + 0.0022310265339910984, + 0.004553302191197872, + 0.002944894367828965, + 0.0015470663784071803, + 0.0034404860343784094, + 0.005792442709207535, + 0.0032023307867348194, + 0.008722544647753239, + 0.002509414218366146, + 0.001912810723297298, + 0.0021626115776598454, + 0.002538020722568035, + 0.001809996087104082, + 0.0017790409037843347, + 0.0016254554502665997, + 0.0019741191063076258, + 0.001753901713527739, + 0.02420959435403347, + 0.313902348279953, + 0.016927478834986687, + 0.3257099986076355, + 0.03140348196029663, + 0.020438913255929947, + 0.05416348576545715, + 0.037815481424331665, + 0.0011060737306252122, + 0.0009414993692189455, + 0.0010390615789219737, + 0.0016165744746103883, + 0.0010400363244116306, + 0.0010392055846750736, + 0.0009278748766519129, + 0.001153507619164884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.2813270092010498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 sterile semiconductor manufacturing facility with a polished, reflective white epoxy floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a multi-jointed robotic arm with polished chrome and white composite surfaces visible extending a vacuum gripper to remove a silicon chip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the silicon chip display iridescent microscopic circuitry that refracts light into a rainbow of colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a technician wearing a spotless, seamless white cleanroom suit with an integrated hood and transparent face shield visible interacting with control interfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle shot and move smoothly forward, then curve left to reveal the technician before continuing forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493161827325821 + ], + "opposite_logprob_tokens": [ + -2.929931640625 + ], + "logprob_reward": -0.05493161827325821, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493161827325821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04156062379479408, + 0.034194473177194595, + 0.028677446767687798, + 0.015166102908551693, + 0.018552355468273163, + 0.02426495961844921, + 0.05088966339826584, + 0.027671409770846367, + 0.7406818270683289, + 0.01834111101925373 + ], + "credit_map": [ + 0.003658697474747896, + 0.0038190060295164585, + 0.007405967451632023, + 0.0053571066819131374, + 0.004780136980116367, + 0.0055008684284985065, + 0.004644467495381832, + 0.006394370924681425, + 0.0029384877998381853, + 0.003640057984739542, + 0.005070449784398079, + 0.004936555866152048, + 0.0037920530885457993, + 0.004376519471406937, + 0.004573867656290531, + 0.004866478033363819, + 0.002573074074462056, + 0.0041548581793904305, + 0.003699609776958823, + 0.0034654419869184494, + 0.004520059563219547, + 0.004782912787050009, + 0.002572103403508663, + 0.0029093835037201643, + 0.0012435178505256772, + 0.0014066598378121853, + 0.0020783613435924053, + 0.002169437473639846, + 0.001514459727331996, + 0.001628500409424305, + 0.0024748549330979586, + 0.0026503107510507107, + 0.002087715780362487, + 0.0021045661997050047, + 0.001922492403537035, + 0.0024082311429083347, + 0.002038676757365465, + 0.0021914306562393904, + 0.0025199423544108868, + 0.0032793001737445593, + 0.002766567515209317, + 0.002274761674925685, + 0.002444509183987975, + 0.0021176638547331095, + 0.002387091750279069, + 0.004282206296920776, + 0.004937394987791777, + 0.0030547608621418476, + 0.007179016247391701, + 0.004617128986865282, + 0.0024415715597569942, + 0.005456184037029743, + 0.008560302667319775, + 0.0047816201113164425, + 0.013912197202444077, + 0.003941641189157963, + 0.003748777322471142, + 0.004832166247069836, + 0.003666243050247431, + 0.002887866459786892, + 0.0028090928681194782, + 0.0029175879899412394, + 0.0034103491343557835, + 0.003399324370548129, + 0.023164838552474976, + 0.27332794666290283, + 0.018981557339429855, + 0.29689276218414307, + 0.02681545354425907, + 0.018292034044861794, + 0.04911082610487938, + 0.0340963751077652, + 0.00147363287396729, + 0.0027317998465150595, + 0.0018226770916953683, + 0.0030780744273215532, + 0.0034536058083176613, + 0.0018083282047882676, + 0.0021029659546911716, + 0.0018700254149734974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.05493161827325821 + } + ], + "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_5/index_tensor([8989, 8989])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_5/index_tensor([8989, 8989])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7df29a5772c555a3d48d2c6f2979c9a16cb0412c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_5/index_tensor([8989, 8989])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "rapid downward tracking shot follows a massive, irregularly shaped, charcoal-black boulder, its surface riddled with sharp angles and deep fissures. Its edges glow an ominous dark red from the high-speed friction of the air. With a sharp whistling sound, it plummets from the gloomy clouds like a meteorite. Below, a quaint, single-story tiled house nestles quietly in a verdant valley, its slate-gray tiled roof shimmering in the diffuse sunlight, a serene and ancient presence. In a fraction of a second, the boulder crashes into the center of the roof with unstoppable, destructive force. Instantly, the fragile tiles shatter in all directions like an explosion, the sturdy wooden beams snap, and a massive cloud of dust, rubble, and sawdust erupts, completely engulfing the previously tranquil rooftop. The entire scene is filled with a sudden, shocking sense of destructive power and dynamism.", + "seed": 713, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_5/index_8989_seed_713_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": -0.019127670116722582, + "vlm_logprob": -0.019127670116722582, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.019127670116722582, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0228415634483099 + ], + "opposite_logprob_tokens": [ + -4.022841453552246 + ], + "logprob_reward": -0.0228415634483099, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0228415634483099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05374899506568909, + 0.21934163570404053, + 0.20542851090431213, + 0.06658971309661865, + 0.03192489221692085, + 0.07170645147562027, + 0.12960979342460632, + 0.07109414041042328, + 0.05944869667291641, + 0.09110717475414276 + ], + "credit_map": [ + 0.005215445999056101, + 0.005195618607103825, + 0.007700040936470032, + 0.006000254303216934, + 0.003962429706007242, + 0.010069849900901318, + 0.009037424810230732, + 0.006567932665348053, + 0.017455046996474266, + 0.0206405371427536, + 0.05048779770731926, + 0.023923231288790703, + 0.018856607377529144, + 0.030027808621525764, + 0.022320834919810295, + 0.0356297604739666, + 0.024162178859114647, + 0.024569274857640266, + 0.037040986120700836, + 0.03043198771774769, + 0.02176508866250515, + 0.029961949214339256, + 0.012512853369116783, + 0.02498418278992176, + 0.007284232880920172, + 0.008846810087561607, + 0.008299953304231167, + 0.00539292860776186, + 0.008814548142254353, + 0.009103459306061268, + 0.007639141287654638, + 0.011208644136786461, + 0.002556356368586421, + 0.006867688614875078, + 0.002911526942625642, + 0.002911773044615984, + 0.00524129718542099, + 0.003828517161309719, + 0.0037595522589981556, + 0.0038481829687952995, + 0.008199392817914486, + 0.007196820806711912, + 0.003556674113497138, + 0.009576155804097652, + 0.01464818324893713, + 0.012778636999428272, + 0.01076007541269064, + 0.004990514367818832, + 0.007082394324243069, + 0.009296299889683723, + 0.008676571771502495, + 0.030965430662035942, + 0.037649236619472504, + 0.019188379868865013, + 0.011003420688211918, + 0.0057480670511722565, + 0.004635517951101065, + 0.006218399852514267, + 0.008835475891828537, + 0.01899617724120617, + 0.007434243801981211, + 0.004391933791339397, + 0.005571218207478523, + 0.015011168085038662, + 0.003958245739340782, + 0.016900258138775826, + 0.002757573500275612, + 0.015209512785077095, + 0.009619650430977345, + 0.002504688687622547, + 0.005136735271662474, + 0.0033620351459831, + 0.007350270636379719, + 0.008986994624137878, + 0.010020154528319836, + 0.02329683117568493, + 0.012458770535886288, + 0.005778566002845764, + 0.009301620535552502, + 0.01391396764665842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.0228415634483099 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601926371455193 + ], + "opposite_logprob_tokens": [ + -3.895601987838745 + ], + "logprob_reward": -0.020601926371455193, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601926371455193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07124751806259155, + 0.1404295116662979, + 0.0790553018450737, + 0.0741669312119484, + 0.03158433362841606, + 0.08405176550149918, + 0.3189561069011688, + 0.0706862211227417, + 0.08886954188346863, + 0.04095275700092316 + ], + "credit_map": [ + 0.005630084779113531, + 0.007033231668174267, + 0.01006719097495079, + 0.007483427878469229, + 0.006020265631377697, + 0.014363085851073265, + 0.01098440121859312, + 0.009665834717452526, + 0.010243777185678482, + 0.01468697376549244, + 0.02565499022603035, + 0.017228640615940094, + 0.016783518716692924, + 0.01977122575044632, + 0.019640833139419556, + 0.016419555991888046, + 0.012061064131557941, + 0.011423428542912006, + 0.010093342512845993, + 0.006552453618496656, + 0.011027587577700615, + 0.010804859921336174, + 0.008007829077541828, + 0.009084738790988922, + 0.008641164749860764, + 0.010497371666133404, + 0.008385543711483479, + 0.007445666007697582, + 0.01009439304471016, + 0.008805993013083935, + 0.01061979029327631, + 0.009677012450993061, + 0.0022742734290659428, + 0.004242731723934412, + 0.0033785973209887743, + 0.005080554634332657, + 0.004418156575411558, + 0.0049691894091665745, + 0.004201095085591078, + 0.0030197370797395706, + 0.009129197336733341, + 0.00868951715528965, + 0.004424504470080137, + 0.011866701766848564, + 0.016864700242877007, + 0.014278206042945385, + 0.013045168481767178, + 0.005753767676651478, + 0.01666112244129181, + 0.022821292281150818, + 0.01567935384809971, + 0.08435583114624023, + 0.09350581467151642, + 0.04407893866300583, + 0.02679448574781418, + 0.015059265308082104, + 0.005934678949415684, + 0.0049302466213703156, + 0.011097243055701256, + 0.018312035128474236, + 0.00494816480204463, + 0.004145875107496977, + 0.00689172837883234, + 0.014426253736019135, + 0.006552602630108595, + 0.022868873551487923, + 0.004067329224199057, + 0.02383156679570675, + 0.012778244912624359, + 0.004255243111401796, + 0.007968656718730927, + 0.006547031458467245, + 0.005654504522681236, + 0.005729734897613525, + 0.007480819709599018, + 0.005508624482899904, + 0.005083914380520582, + 0.00332139665260911, + 0.005003522150218487, + 0.0031702383421361446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.020601926371455193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10177312046289444, + 0.11116435378789902, + 0.07324827462434769, + 0.15942659974098206, + 0.07172548770904541, + 0.06564170122146606, + 0.21440750360488892, + 0.0993771031498909, + 0.04840086027979851, + 0.054835058748722076 + ], + "credit_map": [ + 0.009689215570688248, + 0.008381184190511703, + 0.01710490509867668, + 0.012557211332023144, + 0.006843025796115398, + 0.022490615025162697, + 0.012511481530964375, + 0.012195473536849022, + 0.006980502046644688, + 0.008807291276752949, + 0.018631523475050926, + 0.01590035855770111, + 0.010449240915477276, + 0.022348539903759956, + 0.01203923299908638, + 0.016007648780941963, + 0.010825921781361103, + 0.008609448559582233, + 0.01170277502387762, + 0.005606774240732193, + 0.011810366064310074, + 0.009860317222774029, + 0.006364188622683287, + 0.008468473330140114, + 0.014875265769660473, + 0.01929733157157898, + 0.02462444268167019, + 0.012663206085562706, + 0.01739446260035038, + 0.023165756836533546, + 0.014275847934186459, + 0.03313026949763298, + 0.008396330289542675, + 0.012687785550951958, + 0.0050808340311050415, + 0.014527383260428905, + 0.010604203678667545, + 0.006218988914042711, + 0.008487018756568432, + 0.005722930189222097, + 0.00971437431871891, + 0.008796690963208675, + 0.004537352826446295, + 0.006639248691499233, + 0.011283542960882187, + 0.009824313223361969, + 0.008509325794875622, + 0.006336838006973267, + 0.018332449719309807, + 0.018711568787693977, + 0.026820529252290726, + 0.050187110900878906, + 0.04645179212093353, + 0.020483238622546196, + 0.01944691129028797, + 0.013973867520689964, + 0.011301911436021328, + 0.010874737985432148, + 0.008711736649274826, + 0.023680027574300766, + 0.013422922231256962, + 0.007941133342683315, + 0.007386174984276295, + 0.01605845056474209, + 0.003732278710231185, + 0.007956222631037235, + 0.002749514067545533, + 0.012522726319730282, + 0.006190866231918335, + 0.0038968126755207777, + 0.00710785947740078, + 0.004244576208293438, + 0.005334966350346804, + 0.005839149467647076, + 0.007631038315594196, + 0.013826167210936546, + 0.005458036437630653, + 0.0037783284205943346, + 0.00557919405400753, + 0.007388170342892408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.02058335766196251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351304724812508 + ], + "opposite_logprob_tokens": [ + -3.6513512134552 + ], + "logprob_reward": -0.026351304724812508, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351304724812508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04168574512004852, + 0.03368087857961655, + 0.045955948531627655, + 0.04387073218822479, + 0.06923443078994751, + 0.053039275109767914, + 0.4628998041152954, + 0.1354706734418869, + 0.0411592572927475, + 0.07300318777561188 + ], + "credit_map": [ + 0.003622566582635045, + 0.003069444792345166, + 0.006488090381026268, + 0.0031447047367691994, + 0.0033565228804945946, + 0.011415590532124043, + 0.00456777261570096, + 0.006021049804985523, + 0.0025059920735657215, + 0.00486720772460103, + 0.004646708723157644, + 0.0038274137768894434, + 0.004750480409711599, + 0.004632212687283754, + 0.0049454690888524055, + 0.003505394561216235, + 0.005176296457648277, + 0.005960951559245586, + 0.0066317543387413025, + 0.0076693398877978325, + 0.007674727123230696, + 0.005182034336030483, + 0.00304325926117599, + 0.004617584869265556, + 0.004363494459539652, + 0.004736897069960833, + 0.006904111243784428, + 0.0031378744170069695, + 0.0046606725081801414, + 0.007051011547446251, + 0.0033467980101704597, + 0.00966987106949091, + 0.005740808323025703, + 0.011546319350600243, + 0.006913233082741499, + 0.004835669882595539, + 0.011974666267633438, + 0.009322025813162327, + 0.007850074209272861, + 0.011051639914512634, + 0.00858572218567133, + 0.006804272066801786, + 0.0037866446655243635, + 0.006744238547980785, + 0.007677665911614895, + 0.01031030248850584, + 0.004450873006135225, + 0.004679558333009481, + 0.027164313942193985, + 0.03706744685769081, + 0.022315962240099907, + 0.11638765782117844, + 0.1374378651380539, + 0.06428205966949463, + 0.03767932206392288, + 0.020565226674079895, + 0.015298455953598022, + 0.016444658860564232, + 0.008781800977885723, + 0.022432778030633926, + 0.02022276632487774, + 0.00796316098421812, + 0.022781165316700935, + 0.02154589258134365, + 0.0039010997861623764, + 0.005874210968613625, + 0.004041007254272699, + 0.006961521692574024, + 0.005741973873227835, + 0.003920332994312048, + 0.007253146264702082, + 0.0034659644588828087, + 0.006889316253364086, + 0.00609913794323802, + 0.008233035914599895, + 0.023023806512355804, + 0.006729971617460251, + 0.004239275120198727, + 0.006377798970788717, + 0.0114108482375741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.026351304724812508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260198377072811 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260198377072811, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260198377072811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04866356402635574, + 0.04012744873762131, + 0.054283224046230316, + 0.04468351975083351, + 0.027506139129400253, + 0.044189199805259705, + 0.3571103513240814, + 0.04972610995173454, + 0.23073391616344452, + 0.10297653079032898 + ], + "credit_map": [ + 0.004698619246482849, + 0.005011041183024645, + 0.0071105374954640865, + 0.004385685082525015, + 0.00368597568012774, + 0.00970076210796833, + 0.007570423185825348, + 0.0065005188807845116, + 0.0030706296674907207, + 0.0038063572719693184, + 0.00506474357098341, + 0.004855570383369923, + 0.005483673885464668, + 0.008736889809370041, + 0.004430151078850031, + 0.004679433070123196, + 0.006582348141819239, + 0.006769727449864149, + 0.008470737375319004, + 0.007862002588808537, + 0.008493231609463692, + 0.007039798889309168, + 0.003118113148957491, + 0.0059472606517374516, + 0.004854693543165922, + 0.006442813202738762, + 0.005101285874843597, + 0.0036532345693558455, + 0.007129342295229435, + 0.006586506497114897, + 0.004407733678817749, + 0.00650790985673666, + 0.002644909778609872, + 0.00404553534463048, + 0.0028361021541059017, + 0.0024436558596789837, + 0.0047503081150352955, + 0.004210502840578556, + 0.002978447126224637, + 0.003596679074689746, + 0.005393022205680609, + 0.006103503052145243, + 0.0026330268010497093, + 0.004816356580704451, + 0.008272562175989151, + 0.0072940741665661335, + 0.005455110687762499, + 0.004221541807055473, + 0.021871762350201607, + 0.028936157003045082, + 0.01727626472711563, + 0.09194613248109818, + 0.10037586838006973, + 0.047705624252557755, + 0.030649522319436073, + 0.018349049612879753, + 0.00641770102083683, + 0.005675250664353371, + 0.004079511854797602, + 0.010012278333306313, + 0.006583496928215027, + 0.0038878736086189747, + 0.00502731092274189, + 0.008042685687541962, + 0.020563773810863495, + 0.03830592334270477, + 0.009744314476847649, + 0.037984807044267654, + 0.03795979544520378, + 0.020600544288754463, + 0.04067189618945122, + 0.024902863427996635, + 0.009826489724218845, + 0.010062595829367638, + 0.009600209072232246, + 0.02165963687002659, + 0.020931553095579147, + 0.007718751672655344, + 0.013160514645278454, + 0.010016780346632004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.005260198377072811 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "rapid downward tracking shot follows a massive, irregularly shaped, charcoal-black boulder, its surface riddled with sharp angles and deep fissures. Its edges glow an ominous dark red from the high-speed friction of the air. With a sharp whistling sound, it plummets from the gloomy clouds like a meteorite. Below, a quaint, single-story tiled house nestles quietly in a verdant valley, its slate-gray tiled roof shimmering in the diffuse sunlight, a serene and ancient presence. In a fraction of a second, the boulder crashes into the center of the roof with unstoppable, destructive force. Instantly, the fragile tiles shatter in all directions like an explosion, the sturdy wooden beams snap, and a massive cloud of dust, rubble, and sawdust erupts, completely engulfing the previously tranquil rooftop. The entire scene is filled with a sudden, shocking sense of destructive power and dynamism.", + "seed": 714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_5/index_8989_seed_714_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": -0.018637937121093275, + "vlm_logprob": -0.018637937121093275, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.018637937121093275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010853425599634647 + ], + "opposite_logprob_tokens": [ + -5.010853290557861 + ], + "logprob_reward": -0.010853425599634647, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010853425599634647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10740251839160919, + 0.046360649168491364, + 0.14948341250419617, + 0.15855145454406738, + 0.3095640242099762, + 0.042581673711538315, + 0.03995613753795624, + 0.0666898712515831, + 0.0494375005364418, + 0.029972713440656662 + ], + "credit_map": [ + 0.0061701564118266106, + 0.007854517549276352, + 0.004755210597068071, + 0.009037833660840988, + 0.030320458114147186, + 0.015452702529728413, + 0.015678148716688156, + 0.018133491277694702, + 0.005185692571103573, + 0.0050004227086901665, + 0.005277914926409721, + 0.003850274020805955, + 0.007628289517015219, + 0.007777852471917868, + 0.007021856959909201, + 0.00461835041642189, + 0.013536641374230385, + 0.026433877646923065, + 0.01513728778809309, + 0.017742427065968513, + 0.023009609431028366, + 0.01934320107102394, + 0.018199767917394638, + 0.016080619767308235, + 0.01510479673743248, + 0.012579359114170074, + 0.027088381350040436, + 0.018785225227475166, + 0.012477723881602287, + 0.015560622327029705, + 0.03281765803694725, + 0.02413771115243435, + 0.0412505641579628, + 0.04049840569496155, + 0.052135955542325974, + 0.021860338747501373, + 0.04743914306163788, + 0.05487867444753647, + 0.026225093752145767, + 0.025275859981775284, + 0.006175701972097158, + 0.0036543046589940786, + 0.006783840246498585, + 0.00337848998606205, + 0.002614210359752178, + 0.005736440885812044, + 0.005393669009208679, + 0.008845021016895771, + 0.006957618054002523, + 0.0037939741741865873, + 0.005166544578969479, + 0.0035853488370776176, + 0.003421216504648328, + 0.00472855381667614, + 0.005906582344323397, + 0.006396304350346327, + 0.017175018787384033, + 0.004068909212946892, + 0.003428064053878188, + 0.002807596232742071, + 0.005663950927555561, + 0.023142801597714424, + 0.0027560668531805277, + 0.0076474761590361595, + 0.010583382099866867, + 0.00628271047025919, + 0.00866882037371397, + 0.0039466856978833675, + 0.005811646115034819, + 0.004971210844814777, + 0.005928041413426399, + 0.003245009807869792, + 0.003455908503383398, + 0.003956872038543224, + 0.0034992569126188755, + 0.0034146413672715425, + 0.0038253306411206722, + 0.002953900722786784, + 0.005901664029806852, + 0.0029651408549398184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.010853425599634647 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373740240931511 + ], + "opposite_logprob_tokens": [ + -3.4087374210357666 + ], + "logprob_reward": -0.03373740240931511, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373740240931511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11450333148241043, + 0.0225079245865345, + 0.19126655161380768, + 0.09922578930854797, + 0.4135846793651581, + 0.03718284144997597, + 0.02344442345201969, + 0.036527976393699646, + 0.03642585873603821, + 0.025330625474452972 + ], + "credit_map": [ + 0.0070494613610208035, + 0.0076314108446240425, + 0.005024288780987263, + 0.007839375175535679, + 0.035239268094301224, + 0.01631956733763218, + 0.014259587042033672, + 0.021140366792678833, + 0.0029001159127801657, + 0.0020278592128306627, + 0.004325038753449917, + 0.0023698939476162195, + 0.0024714176543056965, + 0.0031320650596171618, + 0.003118362743407488, + 0.0021631703712046146, + 0.024377522990107536, + 0.02638074941933155, + 0.02191983349621296, + 0.01922294870018959, + 0.022870590910315514, + 0.028788113966584206, + 0.02064872346818447, + 0.027058076113462448, + 0.009287445805966854, + 0.009834840893745422, + 0.016881858929991722, + 0.011436156928539276, + 0.008588222786784172, + 0.012014673091471195, + 0.01590217649936676, + 0.0152804059907794, + 0.05250362306833267, + 0.05575912818312645, + 0.06759994477033615, + 0.02972092479467392, + 0.0660441666841507, + 0.0692877545952797, + 0.03832049295306206, + 0.03434866666793823, + 0.005122334696352482, + 0.0028576981276273727, + 0.007529268506914377, + 0.0027022301219403744, + 0.0033155716955661774, + 0.005101500544697046, + 0.004449428990483284, + 0.0061048055067658424, + 0.0045911772176623344, + 0.0019902940839529037, + 0.0026758122257888317, + 0.0023912943433970213, + 0.002175957430154085, + 0.0025744892191141844, + 0.0027201627381145954, + 0.004325236193835735, + 0.009737773798406124, + 0.0026834094896912575, + 0.002360100159421563, + 0.001725593232549727, + 0.0028460726607590914, + 0.012728693895041943, + 0.001305967103689909, + 0.003140366869047284, + 0.00871940329670906, + 0.003692841622978449, + 0.007681259885430336, + 0.002262664958834648, + 0.0039087506011128426, + 0.004540150053799152, + 0.0019135046750307083, + 0.0037072820123285055, + 0.003004674566909671, + 0.004420843441039324, + 0.00329276523552835, + 0.0022352533414959908, + 0.002776258857920766, + 0.0036565635818988085, + 0.0029036495834589005, + 0.003040616400539875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.03373740240931511 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0142059950158 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.0142059950158, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142059950158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20312272012233734, + 0.02163868397474289, + 0.0822247788310051, + 0.12915286421775818, + 0.232624813914299, + 0.09148547798395157, + 0.06676355749368668, + 0.09500620514154434, + 0.05105126276612282, + 0.026929616928100586 + ], + "credit_map": [ + 0.013105005957186222, + 0.0097520025447011, + 0.005155156832188368, + 0.008532611653208733, + 0.06729547679424286, + 0.032370585948228836, + 0.028595291078090668, + 0.038316600024700165, + 0.0025269831530749798, + 0.002391374669969082, + 0.0034074760042130947, + 0.001935721724294126, + 0.0029643201269209385, + 0.0033953285310417414, + 0.0028595963958650827, + 0.0021578834857791662, + 0.008353703655302525, + 0.014044960960745811, + 0.010073709301650524, + 0.007025306113064289, + 0.00767037458717823, + 0.01208340935409069, + 0.008546349592506886, + 0.014426968060433865, + 0.010209045372903347, + 0.013698011636734009, + 0.020309707149863243, + 0.013468291610479355, + 0.010420478880405426, + 0.016824472695589066, + 0.0167331974953413, + 0.027489660307765007, + 0.030107781291007996, + 0.03034612350165844, + 0.029543276876211166, + 0.01743963174521923, + 0.037409014999866486, + 0.03634319081902504, + 0.02654256485402584, + 0.02489321492612362, + 0.01315190177410841, + 0.011146124452352524, + 0.014408084563910961, + 0.008449244312942028, + 0.008513640612363815, + 0.010845856741070747, + 0.012846349738538265, + 0.012124280445277691, + 0.009390910156071186, + 0.006482848897576332, + 0.009661813266575336, + 0.009711810387670994, + 0.0060461461544036865, + 0.007186830975115299, + 0.009023780934512615, + 0.009259413927793503, + 0.023834822699427605, + 0.0061564999632537365, + 0.005796507932245731, + 0.0036050365306437016, + 0.0077990577556192875, + 0.03268155828118324, + 0.004432928282767534, + 0.010699793696403503, + 0.011598831973969936, + 0.004142013844102621, + 0.009292841888964176, + 0.0033762794919312, + 0.0035304317716509104, + 0.007440309040248394, + 0.003825778141617775, + 0.00784477312117815, + 0.0032731981482356787, + 0.0041374084539711475, + 0.0024904271122068167, + 0.0030417530797421932, + 0.003701769281178713, + 0.003032090375199914, + 0.004874822683632374, + 0.0023781463969498873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0142059950158 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084291152656078 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011084291152656078, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084291152656078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14079764485359192, + 0.05567097291350365, + 0.09089185297489166, + 0.08607622236013412, + 0.05599666014313698, + 0.09837841987609863, + 0.1442224085330963, + 0.11033013463020325, + 0.14438007771968842, + 0.07325556874275208 + ], + "credit_map": [ + 0.006628924049437046, + 0.007897662930190563, + 0.0052695972844958305, + 0.010419709607958794, + 0.04225434735417366, + 0.020054355263710022, + 0.025963423773646355, + 0.02230962924659252, + 0.006084303371608257, + 0.004127540159970522, + 0.010024740360677242, + 0.004962158855050802, + 0.005242431070655584, + 0.00930622685700655, + 0.008990048430860043, + 0.006933525204658508, + 0.007844293490052223, + 0.017163235694169998, + 0.008027535863220692, + 0.013196438550949097, + 0.0120121194049716, + 0.012084527872502804, + 0.008903712034225464, + 0.011659988202154636, + 0.005917564500123262, + 0.007410520222038031, + 0.010480990633368492, + 0.01511223055422306, + 0.006963824853301048, + 0.010448144748806953, + 0.015100378543138504, + 0.014642568305134773, + 0.008119551464915276, + 0.0049150786362588406, + 0.010622910223901272, + 0.005422617774456739, + 0.005300115793943405, + 0.009350549429655075, + 0.006544236559420824, + 0.005721604451537132, + 0.027918649837374687, + 0.011452702805399895, + 0.015359161421656609, + 0.00510438485071063, + 0.008556886576116085, + 0.012096728198230267, + 0.007327641826122999, + 0.010562260635197163, + 0.01868600957095623, + 0.013994901441037655, + 0.010827644728124142, + 0.013133807107806206, + 0.02896822802722454, + 0.013670975342392921, + 0.02729334868490696, + 0.017647497355937958, + 0.017256828024983406, + 0.014384500682353973, + 0.010142610408365726, + 0.01281797792762518, + 0.014168187975883484, + 0.016515614464879036, + 0.008874700404703617, + 0.01616971753537655, + 0.02284415066242218, + 0.019681937992572784, + 0.014067412354052067, + 0.012968378141522408, + 0.028497595340013504, + 0.017677804455161095, + 0.012290483340620995, + 0.016352316364645958, + 0.008833765983581543, + 0.013758617453277111, + 0.008294880390167236, + 0.007084363605827093, + 0.01305423490703106, + 0.007721013855189085, + 0.008113959804177284, + 0.006394735537469387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.011084291152656078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330857142806053 + ], + "opposite_logprob_tokens": [ + -3.773308515548706 + ], + "logprob_reward": -0.02330857142806053, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330857142806053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12289003282785416, + 0.03840964287519455, + 0.07751885056495667, + 0.13302262127399445, + 0.07976990193128586, + 0.09336679428815842, + 0.1000533327460289, + 0.09501763433218002, + 0.16059274971485138, + 0.0993584394454956 + ], + "credit_map": [ + 0.008064444176852703, + 0.008204144425690174, + 0.007195579819381237, + 0.009665577672421932, + 0.025603370741009712, + 0.019919831305742264, + 0.02910250797867775, + 0.015134578570723534, + 0.004173008259385824, + 0.0033013070933520794, + 0.0071999309584498405, + 0.003635762259364128, + 0.0046197338961064816, + 0.0055526080541312695, + 0.006074021570384502, + 0.0038532705511897802, + 0.007063247263431549, + 0.012123586609959602, + 0.007203520275652409, + 0.009949121624231339, + 0.010805067606270313, + 0.010887784883379936, + 0.009455647319555283, + 0.010030880570411682, + 0.014371703378856182, + 0.013175019063055515, + 0.022241150960326195, + 0.019177120178937912, + 0.011711033061146736, + 0.015322072431445122, + 0.019564559683203697, + 0.017459962517023087, + 0.009679263457655907, + 0.006234991829842329, + 0.013770629651844501, + 0.008351088501513004, + 0.010346326045691967, + 0.01112793292850256, + 0.012285664677619934, + 0.00797400251030922, + 0.014593291096389294, + 0.011140946298837662, + 0.012837138026952744, + 0.008145865052938461, + 0.007664155215024948, + 0.012026885524392128, + 0.013694236055016518, + 0.013264277018606663, + 0.012777415104210377, + 0.01146506704390049, + 0.00817256048321724, + 0.011794018559157848, + 0.01265155803412199, + 0.01137644425034523, + 0.018001534044742584, + 0.013814744539558887, + 0.013849184848368168, + 0.01006847433745861, + 0.012551604770123959, + 0.008956816978752613, + 0.01585172675549984, + 0.013427100144326687, + 0.008998843841254711, + 0.011313884519040585, + 0.02889867313206196, + 0.018232041969895363, + 0.027719588950276375, + 0.010199260897934437, + 0.020029574632644653, + 0.015897275879979134, + 0.019378436729311943, + 0.020237913355231285, + 0.011058395728468895, + 0.017277313396334648, + 0.010862358845770359, + 0.008070193231105804, + 0.01577179506421089, + 0.012519127689301968, + 0.013630489818751812, + 0.010168764740228653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.02330857142806053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "rapid downward tracking shot follows a massive, irregularly shaped, charcoal-black boulder, its surface riddled with sharp angles and deep fissures. Its edges glow an ominous dark red from the high-speed friction of the air. With a sharp whistling sound, it plummets from the gloomy clouds like a meteorite. Below, a quaint, single-story tiled house nestles quietly in a verdant valley, its slate-gray tiled roof shimmering in the diffuse sunlight, a serene and ancient presence. In a fraction of a second, the boulder crashes into the center of the roof with unstoppable, destructive force. Instantly, the fragile tiles shatter in all directions like an explosion, the sturdy wooden beams snap, and a massive cloud of dust, rubble, and sawdust erupts, completely engulfing the previously tranquil rooftop. The entire scene is filled with a sudden, shocking sense of destructive power and dynamism.", + "seed": 715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_5/index_8989_seed_715_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": -0.03199871648102999, + "vlm_logprob": -0.03199871648102999, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03199871648102999, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0518299862742424 + ], + "opposite_logprob_tokens": [ + -3.0518300533294678 + ], + "logprob_reward": -0.0518299862742424, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0518299862742424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2828935384750366, + 0.11760393530130386, + 0.0774727538228035, + 0.0911247730255127, + 0.08436373621225357, + 0.03917156904935837, + 0.03253370523452759, + 0.025967251509428024, + 0.20977117121219635, + 0.03909754008054733 + ], + "credit_map": [ + 0.03601096197962761, + 0.031115859746932983, + 0.04735979437828064, + 0.02747376821935177, + 0.029178472235798836, + 0.02749035879969597, + 0.03283348307013512, + 0.0514308325946331, + 0.013969708234071732, + 0.016231007874011993, + 0.015304976142942905, + 0.013194629922509193, + 0.01460304856300354, + 0.010842598043382168, + 0.01659340038895607, + 0.01686456985771656, + 0.007754626218229532, + 0.011636045761406422, + 0.009383152239024639, + 0.007695404347032309, + 0.007636547088623047, + 0.014091228134930134, + 0.010642743669450283, + 0.00863301195204258, + 0.011270794086158276, + 0.011750812642276287, + 0.011228111572563648, + 0.015143905766308308, + 0.010210467502474785, + 0.011520660482347012, + 0.012952997349202633, + 0.007047025952488184, + 0.00960981659591198, + 0.010328363627195358, + 0.019131826236844063, + 0.01033945381641388, + 0.007721272762864828, + 0.008028087206184864, + 0.011446940712630749, + 0.007757974322885275, + 0.005130679812282324, + 0.007327111437916756, + 0.0042626261711120605, + 0.005158562678843737, + 0.005921859294176102, + 0.00404279213398695, + 0.0037898714654147625, + 0.00353806558996439, + 0.0030664231162518263, + 0.003944701980799437, + 0.0045228684321045876, + 0.004313501995056868, + 0.004602887202054262, + 0.004295064136385918, + 0.005215522367507219, + 0.002572734374552965, + 0.0028011437971144915, + 0.005294424016028643, + 0.0036164983175694942, + 0.0030628752429038286, + 0.003426722949370742, + 0.0033290593419224024, + 0.002262721536681056, + 0.0021738074719905853, + 0.018607234582304955, + 0.07738466560840607, + 0.009675281122326851, + 0.009534267708659172, + 0.015528989024460316, + 0.06799864768981934, + 0.004111026879400015, + 0.006931049283593893, + 0.005355819594115019, + 0.0060061807744205, + 0.0031038464512676, + 0.0039504943415522575, + 0.007291877642273903, + 0.004359718877822161, + 0.005115108098834753, + 0.003914498258382082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0518299862742424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373712345957756 + ], + "opposite_logprob_tokens": [ + -3.6513736248016357 + ], + "logprob_reward": -0.026373712345957756, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373712345957756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2243799865245819, + 0.05859583988785744, + 0.08971396088600159, + 0.09085927903652191, + 0.06873621046543121, + 0.03971080109477043, + 0.050174176692962646, + 0.0487443171441555, + 0.2806716859340668, + 0.04841381683945656 + ], + "credit_map": [ + 0.02958572655916214, + 0.029020799323916435, + 0.0295400433242321, + 0.023349344730377197, + 0.023987384513020515, + 0.015514858067035675, + 0.03094257414340973, + 0.04243924096226692, + 0.007099971640855074, + 0.009128818288445473, + 0.0076997303403913975, + 0.007063296157866716, + 0.005963669158518314, + 0.006840345449745655, + 0.007296685129404068, + 0.0075033241882920265, + 0.008574227802455425, + 0.018974341452121735, + 0.010911275632679462, + 0.005197536665946245, + 0.00986668560653925, + 0.013674434274435043, + 0.012017450295388699, + 0.010498005896806717, + 0.011827416718006134, + 0.00975206308066845, + 0.011364307254552841, + 0.017380589619278908, + 0.010133947245776653, + 0.010595383122563362, + 0.01145368255674839, + 0.008351890370249748, + 0.0071128010749816895, + 0.007149550598114729, + 0.013120180927217007, + 0.00904395803809166, + 0.00759999779984355, + 0.006443242076784372, + 0.010783235542476177, + 0.007483244873583317, + 0.004699419252574444, + 0.005878818221390247, + 0.004051392897963524, + 0.006825586780905724, + 0.005217976402491331, + 0.004570154007524252, + 0.004801931791007519, + 0.0036655208095908165, + 0.0035392907448112965, + 0.00815010629594326, + 0.007374352775514126, + 0.005363662727177143, + 0.007581943646073341, + 0.007231836672872305, + 0.007741341833025217, + 0.003191644325852394, + 0.006504826247692108, + 0.010829064063727856, + 0.006578443571925163, + 0.006058542989194393, + 0.00486034108325839, + 0.005688930861651897, + 0.0034781498834490776, + 0.004746018908917904, + 0.02459099143743515, + 0.10528384894132614, + 0.012651398777961731, + 0.011647641658782959, + 0.021062448620796204, + 0.0912591964006424, + 0.005001530516892672, + 0.009174601174890995, + 0.006158065050840378, + 0.007837424986064434, + 0.003464337671175599, + 0.005273994989693165, + 0.008477648720145226, + 0.004606838803738356, + 0.006975521333515644, + 0.005619981326162815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.026373712345957756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37326472997665405, + 0.04521993547677994, + 0.07204011082649231, + 0.05296078696846962, + 0.07970999926328659, + 0.10103052854537964, + 0.105545774102211, + 0.04713459685444832, + 0.09289437532424927, + 0.030199220404028893 + ], + "credit_map": [ + 0.05681461840867996, + 0.03244404494762421, + 0.05094904825091362, + 0.03581469878554344, + 0.027072669938206673, + 0.043680332601070404, + 0.0410277284681797, + 0.0854615792632103, + 0.00644857669249177, + 0.006615246180444956, + 0.006202932447195053, + 0.004765630699694157, + 0.004273212980479002, + 0.005690756719559431, + 0.004580857697874308, + 0.006642721593379974, + 0.00690542021766305, + 0.01132788322865963, + 0.00818333774805069, + 0.010398329235613346, + 0.006273834500461817, + 0.011870232410728931, + 0.008286118507385254, + 0.008794955909252167, + 0.006693883333355188, + 0.0071129463613033295, + 0.007548758294433355, + 0.006754057481884956, + 0.005128084681928158, + 0.010151788592338562, + 0.004398368299007416, + 0.005172903183847666, + 0.009131879545748234, + 0.010627810843288898, + 0.016146734356880188, + 0.009551078081130981, + 0.008837779983878136, + 0.008601666428148746, + 0.009022035636007786, + 0.007791011594235897, + 0.013180301524698734, + 0.014530730433762074, + 0.007599041797220707, + 0.017346592620015144, + 0.015716738998889923, + 0.010117469355463982, + 0.015179536305367947, + 0.0073601193726062775, + 0.009540686383843422, + 0.014636248350143433, + 0.00978272594511509, + 0.014270731247961521, + 0.015614384785294533, + 0.013868720270693302, + 0.013434224762022495, + 0.014398054219782352, + 0.006238811649382114, + 0.012433284893631935, + 0.006332305260002613, + 0.00383567763492465, + 0.005305758677423, + 0.005978561006486416, + 0.0026483964174985886, + 0.004361798986792564, + 0.009658492170274258, + 0.03652380406856537, + 0.004542851354926825, + 0.0029639259446412325, + 0.011165776289999485, + 0.022120676934719086, + 0.001978988526389003, + 0.003939860034734011, + 0.0039406754076480865, + 0.004294497426599264, + 0.002552695805206895, + 0.0019567757844924927, + 0.005871930159628391, + 0.005508280359208584, + 0.0025129353161901236, + 0.0035614315420389175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865137115120888 + ], + "opposite_logprob_tokens": [ + -3.0486514568328857 + ], + "logprob_reward": -0.04865137115120888, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865137115120888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2701787054538727, + 0.07931821793317795, + 0.040317997336387634, + 0.06770924478769302, + 0.0952199250459671, + 0.11275234818458557, + 0.056660380214452744, + 0.10362927615642548, + 0.07157642394304276, + 0.10263744741678238 + ], + "credit_map": [ + 0.0319770947098732, + 0.023491665720939636, + 0.03809555247426033, + 0.02860073745250702, + 0.019669149070978165, + 0.030707480385899544, + 0.02902708575129509, + 0.06860994547605515, + 0.012292396277189255, + 0.008587141521275043, + 0.011562385596334934, + 0.009991979226469994, + 0.008031368255615234, + 0.006961499340832233, + 0.009099164977669716, + 0.012792282737791538, + 0.004035316873341799, + 0.006564212962985039, + 0.0046838982962071896, + 0.005184524226933718, + 0.004094497300684452, + 0.005108846351504326, + 0.006210339721292257, + 0.004436361137777567, + 0.008491715416312218, + 0.010433868505060673, + 0.00736313546076417, + 0.006979922764003277, + 0.006822225172072649, + 0.011581449769437313, + 0.005803811829537153, + 0.010233120061457157, + 0.012080707587301731, + 0.013556521385908127, + 0.02052142843604088, + 0.007275830022990704, + 0.01170498225837946, + 0.01088178064674139, + 0.010582786053419113, + 0.008615884929895401, + 0.015313602983951569, + 0.013823910616338253, + 0.01424974761903286, + 0.01708357036113739, + 0.013101581484079361, + 0.013373278081417084, + 0.014090744778513908, + 0.011715907603502274, + 0.008454815484583378, + 0.0060385060496628284, + 0.0061343214474618435, + 0.006152438931167126, + 0.006491962820291519, + 0.008216289803385735, + 0.008177075535058975, + 0.006994972936809063, + 0.009052693843841553, + 0.010847622528672218, + 0.014008445665240288, + 0.010829533450305462, + 0.010800057090818882, + 0.01882321760058403, + 0.010740206576883793, + 0.018527504056692123, + 0.005652489606291056, + 0.014119718223810196, + 0.005838313605636358, + 0.0035570673644542694, + 0.0072826212272048, + 0.022405942901968956, + 0.0065400125458836555, + 0.006180258002132177, + 0.01385636068880558, + 0.01541534811258316, + 0.006850795354694128, + 0.009110450744628906, + 0.017520975321531296, + 0.010869591496884823, + 0.014431280083954334, + 0.01458264421671629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.04865137115120888 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595036447048187 + ], + "opposite_logprob_tokens": [ + -3.895595073699951 + ], + "logprob_reward": -0.020595036447048187, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595036447048187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17524777352809906, + 0.05942850187420845, + 0.08059661835432053, + 0.049633197486400604, + 0.03759560361504555, + 0.046750932931900024, + 0.07181084901094437, + 0.03355773910880089, + 0.17600077390670776, + 0.2693779766559601 + ], + "credit_map": [ + 0.02253473922610283, + 0.018516961485147476, + 0.023724941536784172, + 0.02122756466269493, + 0.014285801909863949, + 0.019684039056301117, + 0.023686280474066734, + 0.03158746659755707, + 0.008554616943001747, + 0.0062470873817801476, + 0.0076273237355053425, + 0.006782806012779474, + 0.0067775072529911995, + 0.006951051764190197, + 0.0072425249963998795, + 0.00924559123814106, + 0.007949591614305973, + 0.013847794383764267, + 0.0105950478464365, + 0.008706659078598022, + 0.011280279606580734, + 0.008805667981505394, + 0.011718612164258957, + 0.007692975923418999, + 0.0059163630940020084, + 0.005598255433142185, + 0.005230829119682312, + 0.009345100261271, + 0.006011981517076492, + 0.006770140491425991, + 0.004810433369129896, + 0.005950101185590029, + 0.004500257316976786, + 0.005709262564778328, + 0.008229607716202736, + 0.0030156297143548727, + 0.004488404840230942, + 0.004365918226540089, + 0.0033305438701063395, + 0.003955984953790903, + 0.006210493855178356, + 0.006166466977447271, + 0.00518376100808382, + 0.007491001859307289, + 0.006006799638271332, + 0.005514913238584995, + 0.005994576495140791, + 0.004182921256870031, + 0.007142586167901754, + 0.009519447572529316, + 0.009292541071772575, + 0.008525591343641281, + 0.009164274670183659, + 0.009724624454975128, + 0.009766659699380398, + 0.00867513194680214, + 0.004384200554341078, + 0.0061650644056499004, + 0.004328201059252024, + 0.0044471947476267815, + 0.004255045671015978, + 0.004815688356757164, + 0.0025298476684838533, + 0.0026324994396418333, + 0.014860437251627445, + 0.06565491110086441, + 0.008486398495733738, + 0.0076350136660039425, + 0.013478710316121578, + 0.056596968322992325, + 0.003353070467710495, + 0.005935270804911852, + 0.036445099860429764, + 0.04585730656981468, + 0.02671271562576294, + 0.02384893409907818, + 0.047612808644771576, + 0.027777716517448425, + 0.03414887934923172, + 0.02697453834116459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.020595036447048187 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "rapid downward tracking shot follows a massive, irregularly shaped, charcoal-black boulder, its surface riddled with sharp angles and deep fissures. Its edges glow an ominous dark red from the high-speed friction of the air. With a sharp whistling sound, it plummets from the gloomy clouds like a meteorite. Below, a quaint, single-story tiled house nestles quietly in a verdant valley, its slate-gray tiled roof shimmering in the diffuse sunlight, a serene and ancient presence. In a fraction of a second, the boulder crashes into the center of the roof with unstoppable, destructive force. Instantly, the fragile tiles shatter in all directions like an explosion, the sturdy wooden beams snap, and a massive cloud of dust, rubble, and sawdust erupts, completely engulfing the previously tranquil rooftop. The entire scene is filled with a sudden, shocking sense of destructive power and dynamism.", + "seed": 716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_5/index_8989_seed_716_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": -0.018302919529378413, + "vlm_logprob": -0.018302919529378413, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.018302919529378413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011742754839360714 + ], + "opposite_logprob_tokens": [ + -4.88674259185791 + ], + "logprob_reward": -0.011742754839360714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011742754839360714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18160121142864227, + 0.05584617704153061, + 0.1099194884300232, + 0.12474622577428818, + 0.05215620622038841, + 0.1835361123085022, + 0.04918981343507767, + 0.04575996845960617, + 0.0474933423101902, + 0.1497514247894287 + ], + "credit_map": [ + 0.017657645046710968, + 0.020594080910086632, + 0.025898844003677368, + 0.027992190793156624, + 0.023751022294163704, + 0.020561302080750465, + 0.02176656946539879, + 0.023379554972052574, + 0.007566247601062059, + 0.00826108455657959, + 0.007972316816449165, + 0.00699008721858263, + 0.0071440162137150764, + 0.007520470302551985, + 0.0038499131333082914, + 0.006542039569467306, + 0.014078512787818909, + 0.015884144231677055, + 0.01758219301700592, + 0.01704702153801918, + 0.013119718059897423, + 0.012021053582429886, + 0.0109871756285429, + 0.009199663065373898, + 0.015217513777315617, + 0.013338339515030384, + 0.015337659046053886, + 0.011768232099711895, + 0.01684429496526718, + 0.015941336750984192, + 0.02384723164141178, + 0.012451625429093838, + 0.0070545002818107605, + 0.008034532889723778, + 0.004667806904762983, + 0.008449451066553593, + 0.006315756589174271, + 0.008857586421072483, + 0.005842925980687141, + 0.0029336465522646904, + 0.0135916443541646, + 0.02818688377737999, + 0.0245685838162899, + 0.024429311975836754, + 0.024564480409026146, + 0.022615477442741394, + 0.02401849441230297, + 0.021561237052083015, + 0.0032682723831385374, + 0.00714659271761775, + 0.006717815063893795, + 0.010418814606964588, + 0.004740241914987564, + 0.0059564318507909775, + 0.005603432189673185, + 0.005338209215551615, + 0.0034361863508820534, + 0.007490260060876608, + 0.003692532191053033, + 0.007458135951310396, + 0.004418921656906605, + 0.005546363070607185, + 0.008049196563661098, + 0.005668374709784985, + 0.006562810391187668, + 0.008919517509639263, + 0.003461668733507395, + 0.00653626536950469, + 0.006735221482813358, + 0.004509505350142717, + 0.007328014820814133, + 0.0034403398167341948, + 0.015834210440516472, + 0.016594447195529938, + 0.03127231448888779, + 0.017677588388323784, + 0.016720397397875786, + 0.01572704128921032, + 0.017713379114866257, + 0.018212057650089264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.011742754839360714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088063940405846 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011088063940405846, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088063940405846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10153721272945404, + 0.2877322733402252, + 0.09347018599510193, + 0.19989177584648132, + 0.07535921037197113, + 0.049795567989349365, + 0.01972774788737297, + 0.07618309557437897, + 0.035781510174274445, + 0.060521453619003296 + ], + "credit_map": [ + 0.012873025611042976, + 0.009636994451284409, + 0.015542706474661827, + 0.010854008607566357, + 0.010683314874768257, + 0.013997328467667103, + 0.013779397122561932, + 0.01417043711990118, + 0.028578413650393486, + 0.05482684448361397, + 0.03254004940390587, + 0.021777093410491943, + 0.05945985019207001, + 0.03407667204737663, + 0.02345477230846882, + 0.033018577843904495, + 0.014734482392668724, + 0.01487699244171381, + 0.009977954439818859, + 0.0125268017873168, + 0.014329411089420319, + 0.011349927634000778, + 0.00866541638970375, + 0.007009206805378199, + 0.02291608788073063, + 0.025563228875398636, + 0.01996462792158127, + 0.012843728065490723, + 0.029200168326497078, + 0.01631953753530979, + 0.0581582672894001, + 0.01492614671587944, + 0.01205398328602314, + 0.008288897573947906, + 0.005314509384334087, + 0.0058143651112914085, + 0.005651903338730335, + 0.019377900287508965, + 0.009296325035393238, + 0.009561320766806602, + 0.004369980189949274, + 0.007118820212781429, + 0.005654660519212484, + 0.00683890376240015, + 0.005250428803265095, + 0.009091093204915524, + 0.006337723229080439, + 0.005133957602083683, + 0.0013173066545277834, + 0.0026800092309713364, + 0.0039907703176140785, + 0.003265607403591275, + 0.0016049654223024845, + 0.003040471114218235, + 0.0018257871270179749, + 0.0020028301514685154, + 0.006217889487743378, + 0.016119323670864105, + 0.007240369450300932, + 0.011434908024966717, + 0.011396905407309532, + 0.007067318540066481, + 0.011617328971624374, + 0.00508905528113246, + 0.0034125372767448425, + 0.004941823426634073, + 0.003184553235769272, + 0.005472114309668541, + 0.005127430893480778, + 0.00631039310246706, + 0.005253614857792854, + 0.0020790444687008858, + 0.007253487128764391, + 0.0061183106154203415, + 0.011963579803705215, + 0.008427833206951618, + 0.005814495030790567, + 0.005436368752270937, + 0.008405349217355251, + 0.007102031260728836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.011088063940405846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369452804327011 + ], + "opposite_logprob_tokens": [ + -3.4086945056915283 + ], + "logprob_reward": -0.03369452804327011, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369452804327011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1234525591135025, + 0.3001023232936859, + 0.07603387534618378, + 0.10260440409183502, + 0.032785799354314804, + 0.07226775586605072, + 0.03631443530321121, + 0.05631408095359802, + 0.12955661118030548, + 0.07056823372840881 + ], + "credit_map": [ + 0.01142908912152052, + 0.016718951985239983, + 0.01609581522643566, + 0.014781597070395947, + 0.018041832372546196, + 0.01519127655774355, + 0.014373339712619781, + 0.01682065613567829, + 0.03353407233953476, + 0.05687636137008667, + 0.037507832050323486, + 0.023448245599865913, + 0.05730484053492546, + 0.03709213435649872, + 0.02617146447300911, + 0.02816735953092575, + 0.01056907419115305, + 0.012228593230247498, + 0.009368867613375187, + 0.007536380551755428, + 0.012922115623950958, + 0.007920603267848492, + 0.006799560971558094, + 0.008688676171004772, + 0.010293012484908104, + 0.009320469573140144, + 0.01866353675723076, + 0.008132914081215858, + 0.010432709008455276, + 0.01520618237555027, + 0.014891604892909527, + 0.015663979575037956, + 0.0025210059247910976, + 0.005007559899240732, + 0.004018296953290701, + 0.004740446340292692, + 0.004503567237406969, + 0.0032862937077879906, + 0.006045549642294645, + 0.0026630787178874016, + 0.004383751656860113, + 0.009367370046675205, + 0.010041924193501472, + 0.012858634814620018, + 0.006375398952513933, + 0.012376593425869942, + 0.007151789031922817, + 0.00971229374408722, + 0.0027145380154252052, + 0.005309749860316515, + 0.0035525071434676647, + 0.004312390461564064, + 0.004262976348400116, + 0.006885480601340532, + 0.004938709083944559, + 0.004338081926107407, + 0.00478409044444561, + 0.010227343067526817, + 0.007681059185415506, + 0.007296242285519838, + 0.007319912780076265, + 0.006874301936477423, + 0.00677117146551609, + 0.005359956528991461, + 0.02450689487159252, + 0.026025980710983276, + 0.019325710833072662, + 0.012357604689896107, + 0.01808970235288143, + 0.011178146116435528, + 0.009492184035480022, + 0.008580382913351059, + 0.00812851544469595, + 0.006804291624575853, + 0.011060759425163269, + 0.010544967837631702, + 0.007197666447609663, + 0.011339513584971428, + 0.010171188041567802, + 0.005321331787854433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.03369452804327011 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343291625380516 + ], + "opposite_logprob_tokens": [ + -3.65134334564209 + ], + "logprob_reward": -0.026343291625380516, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343291625380516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0634256899356842, + 0.13235192000865936, + 0.02910616062581539, + 0.1536305993795395, + 0.055651675909757614, + 0.24820515513420105, + 0.04919033497571945, + 0.1545678675174713, + 0.054077066481113434, + 0.05979345366358757 + ], + "credit_map": [ + 0.0075993468053638935, + 0.005016324110329151, + 0.011585338041186333, + 0.007662163581699133, + 0.0046958220191299915, + 0.008176613599061966, + 0.01222085952758789, + 0.0064692310988903046, + 0.015509950928390026, + 0.02282572165131569, + 0.016899531707167625, + 0.013411366380751133, + 0.02183787152171135, + 0.015444749966263771, + 0.009855262003839016, + 0.01656746305525303, + 0.005046619568020105, + 0.0028304224833846092, + 0.004164896905422211, + 0.003697579028084874, + 0.003868388244882226, + 0.004142967518419027, + 0.0025395257398486137, + 0.0028157634660601616, + 0.018665576353669167, + 0.011988001875579357, + 0.027952732518315315, + 0.018275147303938866, + 0.01595955714583397, + 0.021486209705471992, + 0.015612253919243813, + 0.023691140115261078, + 0.008954162709414959, + 0.006619513034820557, + 0.004749095533043146, + 0.007257113233208656, + 0.006319527048617601, + 0.009803356602787971, + 0.008367949165403843, + 0.0035809637047350407, + 0.0185959842056036, + 0.03454507514834404, + 0.03710952401161194, + 0.04784657806158066, + 0.028717007488012314, + 0.028407778590917587, + 0.02718147449195385, + 0.025801753625273705, + 0.003622023155912757, + 0.007774176076054573, + 0.005219284910708666, + 0.007867449894547462, + 0.00998381245881319, + 0.006511468440294266, + 0.0047213234938681126, + 0.0034908002708107233, + 0.013025962747633457, + 0.031805284321308136, + 0.014583280310034752, + 0.02873358502984047, + 0.020681781694293022, + 0.01565474085509777, + 0.017713718116283417, + 0.012369520962238312, + 0.010345065034925938, + 0.008035174570977688, + 0.005487370770424604, + 0.0062347035855054855, + 0.007415891159325838, + 0.005609087646007538, + 0.005583799444139004, + 0.005365975201129913, + 0.006219281814992428, + 0.00595935620367527, + 0.011216477490961552, + 0.007947999984025955, + 0.005324821453541517, + 0.008517017588019371, + 0.0045870039612054825, + 0.01002150122076273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.026343291625380516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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, irregularly shaped, charcoal-black boulder with sharp angles and deep fissures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boulder's edges glow an ominous dark red due to high-speed friction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera executing a rapid downward tracking shot following the boulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boulder crash into the center of a quaint, single-story house with a slate-gray tiled roof in a verdant valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the impact cause the tiles to shatter, wooden beams to snap, and a massive cloud of dust, rubble, and sawdust to erupt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645959198474884 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008645959198474884, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645959198474884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12234045565128326, + 0.047655295580625534, + 0.027521418407559395, + 0.07322109490633011, + 0.09699546545743942, + 0.07397793978452682, + 0.07677549868822098, + 0.35749873518943787, + 0.04930916428565979, + 0.07470499724149704 + ], + "credit_map": [ + 0.011535167694091797, + 0.011979415081441402, + 0.017195263877511024, + 0.018800023943185806, + 0.01224239356815815, + 0.014238808304071426, + 0.019732311367988586, + 0.016617072746157646, + 0.00538366986438632, + 0.007733568083494902, + 0.00733558926731348, + 0.004469108302146196, + 0.007653174921870232, + 0.005780368112027645, + 0.003378436202183366, + 0.005921381991356611, + 0.004414848051965237, + 0.0043773651123046875, + 0.0031764013692736626, + 0.0023217243142426014, + 0.004500884562730789, + 0.0027693272568285465, + 0.002676896983757615, + 0.003283973317593336, + 0.0076626986265182495, + 0.0076316604390740395, + 0.011342438869178295, + 0.0054220701567828655, + 0.008573372848331928, + 0.006450348999351263, + 0.01832127943634987, + 0.007817227393388748, + 0.028217002749443054, + 0.0063977669924497604, + 0.005369103047996759, + 0.006409323308616877, + 0.006263135001063347, + 0.03396446630358696, + 0.005250226706266403, + 0.005124440882354975, + 0.005552317947149277, + 0.010311098769307137, + 0.008573413826525211, + 0.010434169322252274, + 0.011451675556600094, + 0.007631361018866301, + 0.01208643801510334, + 0.007937463000416756, + 0.005382864736020565, + 0.014105771668255329, + 0.010196836665272713, + 0.013962010852992535, + 0.008064433932304382, + 0.009793480858206749, + 0.007240837439894676, + 0.00802926067262888, + 0.02969191037118435, + 0.07498820871114731, + 0.034562207758426666, + 0.06974413990974426, + 0.04901346564292908, + 0.028460128232836723, + 0.04709900543093681, + 0.02393965795636177, + 0.00767119787633419, + 0.008929404430091381, + 0.005628873128443956, + 0.0046011763624846935, + 0.00944328773766756, + 0.004822415299713612, + 0.005266057327389717, + 0.002946750493720174, + 0.009318315424025059, + 0.007696068380028009, + 0.010756914503872395, + 0.013194732367992401, + 0.0076603274792432785, + 0.008522254414856434, + 0.011505791917443275, + 0.006050593685358763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.008645959198474884 + } + ], + "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_6/index_tensor([5653, 5653])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_tensor([5653, 5653])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e788950359c9070c704cada758002871a24f9ea --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_tensor([5653, 5653])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, right-panning, wide-angle aerial shot, capturing the unmistakable aesthetic of Hayao Miyazaki's animated films, slowly sweeps across a vast, lush valley. The soft afternoon sun bathes everything in a warm golden hue, casting a soft glow over the rolling hills. A few large, fluffy, sharply defined cotton candy clouds drift across the sky. Amidst this magnificent natural landscape, a lone explorer, carrying a large, vintage travel bag, walks along a narrow dirt path winding through the meadow. Dressed in a durable beige canvas jacket, dark trousers, and sturdy hiking boots, the explorer's massive leather and thick canvas travel bag, overflowing with belongings, swayed slightly with his steady, deliberate strides. A metal water bottle and a rolled-up sleeping bag dangle from its side. The camera follows him, his back appears small and determined in the vast world, as he walks step by step towards the unknown fantasy world shrouded in mist and with strange outlines on the distant horizon. The whole picture is full of tranquil poetry and infinite yearning for an unknown journey, creating a peaceful, lonely and hopeful atmosphere.", + "seed": 15346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_5653_seed_15346_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": -0.059252917859703304, + "vlm_logprob": -0.059252917859703304, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.059252917859703304, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008966065011918545 + ], + "opposite_logprob_tokens": [ + -4.883965969085693 + ], + "logprob_reward": -0.008966065011918545, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008966065011918545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15409021079540253, + 0.07174426317214966, + 0.08878073841333389, + 0.12532705068588257, + 0.16183792054653168, + 0.19061151146888733, + 0.03114250674843788, + 0.041765566915273666, + 0.06205017492175102, + 0.0726500004529953 + ], + "credit_map": [ + 0.011340711265802383, + 0.010415381751954556, + 0.03875545412302017, + 0.015310206450521946, + 0.01386355422437191, + 0.010405651293694973, + 0.014563154429197311, + 0.03943610191345215, + 0.009022226557135582, + 0.006415148731321096, + 0.01772012561559677, + 0.008339762687683105, + 0.010206907987594604, + 0.0041817715391516685, + 0.007705374155193567, + 0.008152947761118412, + 0.012204380705952644, + 0.006730820517987013, + 0.015193765982985497, + 0.011830626986920834, + 0.011021627113223076, + 0.0049285381101071835, + 0.016578536480665207, + 0.010292449034750462, + 0.017278406769037247, + 0.01365808118134737, + 0.019549839198589325, + 0.018199386075139046, + 0.008843761868774891, + 0.01645493134856224, + 0.013149207457900047, + 0.0181934405118227, + 0.02740112505853176, + 0.014316968619823456, + 0.019512400031089783, + 0.017757229506969452, + 0.017859766259789467, + 0.02984684333205223, + 0.01430502999573946, + 0.020838551223278046, + 0.023009900003671646, + 0.02646835893392563, + 0.027245404198765755, + 0.02312394790351391, + 0.029590284451842308, + 0.018202725797891617, + 0.019531894475221634, + 0.023438995704054832, + 0.002859798725694418, + 0.003463875036686659, + 0.0029624635353684425, + 0.005068385973572731, + 0.00668344646692276, + 0.0026697360444813967, + 0.0036082982551306486, + 0.0038265029434114695, + 0.0034113044384866953, + 0.004486497957259417, + 0.006106047425419092, + 0.004513371270149946, + 0.006992443930357695, + 0.004524938762187958, + 0.00552293611690402, + 0.0062080263160169125, + 0.009831101633608341, + 0.0034565345849841833, + 0.02230694517493248, + 0.005978004075586796, + 0.002817377680912614, + 0.002366910222917795, + 0.003936519380658865, + 0.011356780305504799, + 0.007393806707113981, + 0.008755707181990147, + 0.015297495760023594, + 0.006924679037183523, + 0.011314972303807735, + 0.004866121802479029, + 0.008688739500939846, + 0.00940848607569933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.008966065011918545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036216876469552517 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036216876469552517, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036216876469552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306154876947403, + 0.1297719031572342, + 0.09119602292776108, + 0.06143473461270332, + 0.04521147534251213, + 0.029537346214056015, + 0.05095784366130829, + 0.06333841383457184, + 0.353301078081131, + 0.04463568702340126 + ], + "credit_map": [ + 0.014650349505245686, + 0.015520910732448101, + 0.025541480630636215, + 0.01566603220999241, + 0.0185884777456522, + 0.013773687183856964, + 0.010331680998206139, + 0.016542868688702583, + 0.01740550622344017, + 0.02069145068526268, + 0.014127744361758232, + 0.01729011908173561, + 0.012966293841600418, + 0.02090957947075367, + 0.009982527233660221, + 0.016398677602410316, + 0.015624667517840862, + 0.00824710913002491, + 0.018709424883127213, + 0.009897496551275253, + 0.011062883771955967, + 0.005865347571671009, + 0.011459587141871452, + 0.01032950822263956, + 0.005963899660855532, + 0.005234630312770605, + 0.009263790212571621, + 0.00825764611363411, + 0.009624664671719074, + 0.004868243355304003, + 0.010433605872094631, + 0.007788251619786024, + 0.007286240812391043, + 0.005507742520421743, + 0.005561325699090958, + 0.004992515314370394, + 0.0050892699509859085, + 0.007504479493945837, + 0.0031744292937219143, + 0.006095472257584333, + 0.0028058120515197515, + 0.0022856879513710737, + 0.005051915999501944, + 0.004131637513637543, + 0.002540667774155736, + 0.0018625168595463037, + 0.004133076872676611, + 0.00672603165730834, + 0.006375808268785477, + 0.006791362538933754, + 0.006158787757158279, + 0.007056667469441891, + 0.005628347862511873, + 0.0043589891865849495, + 0.004912677686661482, + 0.009675201959908009, + 0.00884440541267395, + 0.005098027642816305, + 0.02155141346156597, + 0.007128672208636999, + 0.003857603296637535, + 0.002863486995920539, + 0.004116948693990707, + 0.009877856820821762, + 0.05285615846514702, + 0.020385276526212692, + 0.13385125994682312, + 0.03269475698471069, + 0.009542743675410748, + 0.014461256563663483, + 0.015215500257909298, + 0.0742940828204155, + 0.0055964491330087185, + 0.004140836652368307, + 0.011588877066969872, + 0.004820562433451414, + 0.004453111439943314, + 0.0023947132285684347, + 0.0049729123711586, + 0.0066682216711342335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0036216876469552517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2978302538394928, + 0.18183661997318268, + 0.05247750133275986, + 0.05020812526345253, + 0.05457212030887604, + 0.030592525377869606, + 0.04820557311177254, + 0.04948144406080246, + 0.15237218141555786, + 0.08242364972829819 + ], + "credit_map": [ + 0.02962404489517212, + 0.019295472651720047, + 0.07653488218784332, + 0.02162603661417961, + 0.025567054748535156, + 0.014383000321686268, + 0.017089443281292915, + 0.09371030330657959, + 0.012921861372888088, + 0.013329271227121353, + 0.025693677365779877, + 0.023932551965117455, + 0.033578723669052124, + 0.011090775951743126, + 0.0350663997232914, + 0.026223359629511833, + 0.005392876919358969, + 0.0044036428444087505, + 0.00660356180742383, + 0.008043077774345875, + 0.008413542993366718, + 0.0033370794262737036, + 0.009689382277429104, + 0.006594336125999689, + 0.005094201769679785, + 0.0041018384508788586, + 0.007187787909060717, + 0.005476361606270075, + 0.009519773535430431, + 0.005260526202619076, + 0.007401284761726856, + 0.0061663491651415825, + 0.014520218595862389, + 0.004410001914948225, + 0.003177112899720669, + 0.004506063647568226, + 0.004163088742643595, + 0.016177743673324585, + 0.0041942535899579525, + 0.0034236405044794083, + 0.003322548931464553, + 0.0027270789723843336, + 0.004614759236574173, + 0.004048164468258619, + 0.004900366999208927, + 0.0022345692850649357, + 0.004771627485752106, + 0.003973411396145821, + 0.003867215709760785, + 0.00485383253544569, + 0.006293483078479767, + 0.006681664381176233, + 0.009603578597307205, + 0.00334529229439795, + 0.0064334264025092125, + 0.007127078715711832, + 0.005777460522949696, + 0.0056660533882677555, + 0.0069877710193395615, + 0.00908805150538683, + 0.005610230378806591, + 0.0034281425178050995, + 0.005992707796394825, + 0.006931025069206953, + 0.023399336263537407, + 0.008999300189316273, + 0.0585106797516346, + 0.014738322235643864, + 0.004125770181417465, + 0.004960210062563419, + 0.007409324403852224, + 0.0302292313426733, + 0.010547787882387638, + 0.008959859609603882, + 0.02089119516313076, + 0.006986085791140795, + 0.009977203793823719, + 0.005508433561772108, + 0.009118950925767422, + 0.010434135794639587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "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\": \"Is the shot a wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345033198595047 + ], + "opposite_logprob_tokens": [ + -3.6513450145721436 + ], + "logprob_reward": -0.026345033198595047, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345033198595047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4152281582355499, + 0.06075518950819969, + 0.06226751580834389, + 0.1273607313632965, + 0.09233863651752472, + 0.06168482080101967, + 0.058248698711395264, + 0.040173597633838654, + 0.04161159321665764, + 0.04033103957772255 + ], + "credit_map": [ + 0.04835138097405434, + 0.03163721412420273, + 0.08293206244707108, + 0.03442612662911415, + 0.05715871974825859, + 0.023894386366009712, + 0.033678580075502396, + 0.10314970463514328, + 0.005105033982545137, + 0.007262023165822029, + 0.011417049914598465, + 0.008227108046412468, + 0.007113478612154722, + 0.005692319944500923, + 0.008938082493841648, + 0.007000090088695288, + 0.007574154529720545, + 0.006275427993386984, + 0.006307193543761969, + 0.011119990609586239, + 0.008168316446244717, + 0.005785252898931503, + 0.009430174715816975, + 0.007607006933540106, + 0.01575653627514839, + 0.009444359689950943, + 0.01387752965092659, + 0.019007200375199318, + 0.014583401381969452, + 0.017171405255794525, + 0.023120207712054253, + 0.014400097541511059, + 0.014255791902542114, + 0.006884708069264889, + 0.008447970263659954, + 0.011180190369486809, + 0.012756421230733395, + 0.01897595077753067, + 0.010404183529317379, + 0.009433413855731487, + 0.005690705496817827, + 0.004209305625408888, + 0.010522942058742046, + 0.0079388078302145, + 0.00504775857552886, + 0.004484121222048998, + 0.010253772139549255, + 0.013537407852709293, + 0.006018053740262985, + 0.003842900739982724, + 0.011885566636919975, + 0.008895231410861015, + 0.010425305925309658, + 0.0037670480087399483, + 0.006435243412852287, + 0.006979350466281176, + 0.003371058264747262, + 0.0036076977849006653, + 0.0064623658545315266, + 0.0062801893800497055, + 0.004295610822737217, + 0.0029734193813055754, + 0.004930634051561356, + 0.008252623490989208, + 0.005159140098839998, + 0.005775775294750929, + 0.004517177119851112, + 0.005154092330485582, + 0.0033908127807080746, + 0.004535356070846319, + 0.0038330822717398405, + 0.009246157482266426, + 0.002703102771192789, + 0.002768500940874219, + 0.004946105647832155, + 0.006771453190594912, + 0.006376137491315603, + 0.0023139221593737602, + 0.007837592624127865, + 0.006614224053919315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.026345033198595047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520802617073059 + ], + "opposite_logprob_tokens": [ + -1.5020802021026611 + ], + "logprob_reward": -0.2520802617073059, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520802617073059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2248702347278595, + 0.18478527665138245, + 0.07754746079444885, + 0.07199173420667648, + 0.07440600544214249, + 0.06889180839061737, + 0.036220941692590714, + 0.07930779457092285, + 0.15376895666122437, + 0.02820979245007038 + ], + "credit_map": [ + 0.035568274557590485, + 0.018034761771559715, + 0.03677120432257652, + 0.024433309212327003, + 0.0233269352465868, + 0.013781384564936161, + 0.02072022296488285, + 0.05223413184285164, + 0.016630083322525024, + 0.013025714084506035, + 0.029726432636380196, + 0.021645378321409225, + 0.030484335497021675, + 0.009582390077412128, + 0.037313323467969894, + 0.026377610862255096, + 0.010214692912995815, + 0.007392368279397488, + 0.00738748162984848, + 0.011149254627525806, + 0.01098471600562334, + 0.008673129603266716, + 0.012088696472346783, + 0.009657119400799274, + 0.009822019375860691, + 0.0054574087262153625, + 0.006990167312324047, + 0.011429916135966778, + 0.008437585085630417, + 0.009161736816167831, + 0.012825838290154934, + 0.007867062464356422, + 0.00896799098700285, + 0.0078898835927248, + 0.008635375648736954, + 0.008481214754283428, + 0.013836289756000042, + 0.009513591416180134, + 0.008679223246872425, + 0.008402437902987003, + 0.009584554471075535, + 0.006723994389176369, + 0.0121360057964921, + 0.008413735777139664, + 0.007648763246834278, + 0.005230269394814968, + 0.007890859618782997, + 0.011263619177043438, + 0.0033589182421565056, + 0.004049958195537329, + 0.004050208255648613, + 0.004088120069354773, + 0.007808359805494547, + 0.0030314901378005743, + 0.005844505503773689, + 0.003989379853010178, + 0.006667078472673893, + 0.007166349794715643, + 0.015846505761146545, + 0.008427662774920464, + 0.013864089734852314, + 0.008652977645397186, + 0.008577141910791397, + 0.010105988942086697, + 0.019186660647392273, + 0.007550157140940428, + 0.05735498666763306, + 0.015381387434899807, + 0.005914332810789347, + 0.005247667897492647, + 0.009093162603676319, + 0.03404058888554573, + 0.00303047988563776, + 0.004041542299091816, + 0.003632067237049341, + 0.004516706336289644, + 0.003278559073805809, + 0.0026117642410099506, + 0.0036913391668349504, + 0.0034073353745043278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.2520802617073059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-panning, wide-angle aerial shot, capturing the unmistakable aesthetic of Hayao Miyazaki's animated films, slowly sweeps across a vast, lush valley. The soft afternoon sun bathes everything in a warm golden hue, casting a soft glow over the rolling hills. A few large, fluffy, sharply defined cotton candy clouds drift across the sky. Amidst this magnificent natural landscape, a lone explorer, carrying a large, vintage travel bag, walks along a narrow dirt path winding through the meadow. Dressed in a durable beige canvas jacket, dark trousers, and sturdy hiking boots, the explorer's massive leather and thick canvas travel bag, overflowing with belongings, swayed slightly with his steady, deliberate strides. A metal water bottle and a rolled-up sleeping bag dangle from its side. The camera follows him, his back appears small and determined in the vast world, as he walks step by step towards the unknown fantasy world shrouded in mist and with strange outlines on the distant horizon. The whole picture is full of tranquil poetry and infinite yearning for an unknown journey, creating a peaceful, lonely and hopeful atmosphere.", + "seed": 15347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_5653_seed_15347_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": -0.044052335899323225, + "vlm_logprob": -0.044052335899323225, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.044052335899323225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014279799535870552 + ], + "opposite_logprob_tokens": [ + -4.514279842376709 + ], + "logprob_reward": -0.014279799535870552, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014279799535870552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2925138771533966, + 0.04658069461584091, + 0.13538533449172974, + 0.05929448455572128, + 0.05552768334746361, + 0.11896658688783646, + 0.07728308439254761, + 0.07399030774831772, + 0.09439624100923538, + 0.04606165364384651 + ], + "credit_map": [ + 0.05628131330013275, + 0.05015399307012558, + 0.030178489163517952, + 0.036595262587070465, + 0.03907831758260727, + 0.02616913989186287, + 0.036192797124385834, + 0.017864594236016273, + 0.004813737701624632, + 0.004915881436318159, + 0.006112921983003616, + 0.007495671510696411, + 0.0044346703216433525, + 0.005860215052962303, + 0.006515938322991133, + 0.0064316620118916035, + 0.023076599463820457, + 0.017592784017324448, + 0.013503898866474628, + 0.02925996296107769, + 0.013022296130657196, + 0.012468782253563404, + 0.01092776469886303, + 0.015533261001110077, + 0.008557242341339588, + 0.009479825384914875, + 0.006061985623091459, + 0.009826126508414745, + 0.009054218418896198, + 0.005703727249056101, + 0.005414519924670458, + 0.00519684748724103, + 0.009540986269712448, + 0.008126857690513134, + 0.006530323997139931, + 0.00832382496446371, + 0.006696208380162716, + 0.005908783990889788, + 0.0050137704238295555, + 0.005386933218687773, + 0.022451341152191162, + 0.02150854840874672, + 0.01378678996115923, + 0.008852319791913033, + 0.012533014640212059, + 0.01663069985806942, + 0.007872309535741806, + 0.015331568196415901, + 0.008564694784581661, + 0.008330228738486767, + 0.01720287650823593, + 0.008621744811534882, + 0.011188127100467682, + 0.0057232677936553955, + 0.007697232998907566, + 0.00995491910725832, + 0.009726101532578468, + 0.013534512370824814, + 0.007896962575614452, + 0.010447130538523197, + 0.009736831299960613, + 0.006562660913914442, + 0.008015814237296581, + 0.008070302195847034, + 0.011968068778514862, + 0.016127081587910652, + 0.012966098263859749, + 0.013936023227870464, + 0.008284751325845718, + 0.011201470158994198, + 0.007865303196012974, + 0.01204744540154934, + 0.005819268524646759, + 0.00874332059174776, + 0.005788525566458702, + 0.006337807979434729, + 0.0059493412263691425, + 0.004019466694444418, + 0.00444750627502799, + 0.004956423304975033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.014279799535870552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006753476802259684 + ], + "opposite_logprob_tokens": [ + -5.006753444671631 + ], + "logprob_reward": -0.006753476802259684, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006753476802259684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1657867729663849, + 0.06671496480703354, + 0.23164531588554382, + 0.02693024091422558, + 0.27503225207328796, + 0.04383028671145439, + 0.04212399572134018, + 0.08960965275764465, + 0.036004625260829926, + 0.02232193574309349 + ], + "credit_map": [ + 0.034034982323646545, + 0.02973860129714012, + 0.018606210127472878, + 0.01207356434315443, + 0.031230488792061806, + 0.013170088641345501, + 0.017212945967912674, + 0.009719884023070335, + 0.009547075256705284, + 0.007886956445872784, + 0.009186963550746441, + 0.012181184254586697, + 0.009047102183103561, + 0.006974157877266407, + 0.0060984790325164795, + 0.005793046671897173, + 0.041878145188093185, + 0.030712416395545006, + 0.025431323796510696, + 0.04512957111001015, + 0.02005961909890175, + 0.024643605574965477, + 0.0176281426101923, + 0.026162486523389816, + 0.00413311505690217, + 0.0036260164342820644, + 0.0031384059693664312, + 0.004178712144494057, + 0.0038096990901976824, + 0.0028063629288226366, + 0.002491680672392249, + 0.0027462472207844257, + 0.06922248005867004, + 0.05911172181367874, + 0.023897923529148102, + 0.02950606308877468, + 0.032480936497449875, + 0.026289060711860657, + 0.014606261625885963, + 0.01991778053343296, + 0.005977189633995295, + 0.00786175113171339, + 0.00542484363541007, + 0.003563932841643691, + 0.0055529107339680195, + 0.0067889210768043995, + 0.0039777327328920364, + 0.004683002829551697, + 0.006314206402748823, + 0.008936007507145405, + 0.0029806026723235846, + 0.004409394226968288, + 0.006766784004867077, + 0.003770291106775403, + 0.005392892751842737, + 0.003553814720362425, + 0.010602828115224838, + 0.00988971721380949, + 0.010674191638827324, + 0.014019502326846123, + 0.013776992447674274, + 0.010514182969927788, + 0.008428439497947693, + 0.011703801341354847, + 0.004836292006075382, + 0.0077421776950359344, + 0.0036272399593144655, + 0.004980134777724743, + 0.0050179604440927505, + 0.003409317461773753, + 0.0033324372489005327, + 0.003059064270928502, + 0.0031070359982550144, + 0.004213239531964064, + 0.0035256773699074984, + 0.002409560838714242, + 0.0024973773397505283, + 0.0016637027729302645, + 0.0017784455558285117, + 0.0031268952880054712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.006753476802259684 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16747520864009857, + 0.07884103059768677, + 0.13234259188175201, + 0.11027580499649048, + 0.20680762827396393, + 0.049787599593400955, + 0.06209760531783104, + 0.06369335949420929, + 0.06097869947552681, + 0.06770047545433044 + ], + "credit_map": [ + 0.0209515281021595, + 0.02745644748210907, + 0.020203780382871628, + 0.01458731759339571, + 0.027927206829190254, + 0.022733572870492935, + 0.019577546045184135, + 0.014037806540727615, + 0.005969424266368151, + 0.007635470014065504, + 0.010566865094006062, + 0.010122930631041527, + 0.007123037241399288, + 0.016840070486068726, + 0.00876014493405819, + 0.011823087930679321, + 0.01895967125892639, + 0.017538079991936684, + 0.009458773769438267, + 0.01742522604763508, + 0.028733404353260994, + 0.013866079039871693, + 0.014988826587796211, + 0.01137253176420927, + 0.017913341522216797, + 0.01969967968761921, + 0.009875151328742504, + 0.016382990404963493, + 0.015370494686067104, + 0.009228130802512169, + 0.012883994728326797, + 0.00892200879752636, + 0.05955488234758377, + 0.055807001888751984, + 0.01124198455363512, + 0.013907641172409058, + 0.024169091135263443, + 0.018020836636424065, + 0.011900615878403187, + 0.012205565348267555, + 0.007609950844198465, + 0.007539120968431234, + 0.00713367247954011, + 0.005105425138026476, + 0.006567762698978186, + 0.004435371607542038, + 0.0051497784443199635, + 0.006246522068977356, + 0.007853906601667404, + 0.008084621280431747, + 0.005371744278818369, + 0.007946069352328777, + 0.008637852966785431, + 0.0049483743496239185, + 0.012084315530955791, + 0.007170718163251877, + 0.008489483967423439, + 0.009585507214069366, + 0.005422103684395552, + 0.009356053546071053, + 0.010684222914278507, + 0.006224603392183781, + 0.008147277869284153, + 0.005784109700471163, + 0.007406833581626415, + 0.009870227426290512, + 0.006657040677964687, + 0.008039717562496662, + 0.010567782446742058, + 0.005560039542615414, + 0.006752072833478451, + 0.006124987732619047, + 0.007188522722572088, + 0.010228612460196018, + 0.009813112206757069, + 0.009230662137269974, + 0.007126962766051292, + 0.005463546607643366, + 0.01013409998267889, + 0.008514957502484322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.012537590228021145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1645449995994568, + 0.07755325734615326, + 0.07992830127477646, + 0.049155909568071365, + 0.22797006368637085, + 0.1094340980052948, + 0.07523752748966217, + 0.04943254962563515, + 0.13437871634960175, + 0.03236457705497742 + ], + "credit_map": [ + 0.019452646374702454, + 0.02448843978345394, + 0.022858204320073128, + 0.018333567306399345, + 0.028688820078969002, + 0.015157593414187431, + 0.02192680910229683, + 0.013638917356729507, + 0.005558498203754425, + 0.007913701236248016, + 0.011979471892118454, + 0.015596659854054451, + 0.006710551213473082, + 0.013172932900488377, + 0.007839033380150795, + 0.008782410994172096, + 0.009510899893939495, + 0.010620709508657455, + 0.0065889242105185986, + 0.007840237580239773, + 0.018200350925326347, + 0.008593234233558178, + 0.010714033618569374, + 0.0078599127009511, + 0.005682614166289568, + 0.0056452141143381596, + 0.004842796362936497, + 0.009266002103686333, + 0.006014280952513218, + 0.004269047174602747, + 0.007369485683739185, + 0.0060664634220302105, + 0.06305376440286636, + 0.05894869565963745, + 0.013839758932590485, + 0.019105106592178345, + 0.030569100752472878, + 0.01870221458375454, + 0.010427230969071388, + 0.013324200175702572, + 0.020902786403894424, + 0.019423753023147583, + 0.006846665404736996, + 0.0073704649694263935, + 0.024074256420135498, + 0.006185614038258791, + 0.016398340463638306, + 0.008232209831476212, + 0.01005550567060709, + 0.009253693744540215, + 0.005497860722243786, + 0.010181239806115627, + 0.016245990991592407, + 0.004275895655155182, + 0.011754186823964119, + 0.00797315314412117, + 0.006649498362094164, + 0.00765155628323555, + 0.006293177604675293, + 0.0050750612281262875, + 0.006887214258313179, + 0.007150005549192429, + 0.004137658979743719, + 0.005588379688560963, + 0.018810823559761047, + 0.023005662485957146, + 0.01174703799188137, + 0.013473490253090858, + 0.026267239823937416, + 0.012722593732178211, + 0.014337701722979546, + 0.014014172367751598, + 0.004670337773859501, + 0.005630874540656805, + 0.0033017019741237164, + 0.0036016397643834352, + 0.0052649700082838535, + 0.0031258531380444765, + 0.002814769046381116, + 0.0039544301107525826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.026350026950240135 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603407859802246 + ], + "opposite_logprob_tokens": [ + -1.9103407859802246 + ], + "logprob_reward": -0.1603407859802246, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603407859802246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21657828986644745, + 0.05013992637395859, + 0.12338920682668686, + 0.10880232602357864, + 0.11327893286943436, + 0.0858272835612297, + 0.07519970834255219, + 0.11925040930509567, + 0.07963457703590393, + 0.02789933606982231 + ], + "credit_map": [ + 0.024062225595116615, + 0.03111625276505947, + 0.028589380905032158, + 0.020338719710707664, + 0.030113741755485535, + 0.03736766427755356, + 0.02373052015900612, + 0.02125977911055088, + 0.005665520206093788, + 0.005809704307466745, + 0.004785980563610792, + 0.007593696471303701, + 0.004986763000488281, + 0.008216551505029202, + 0.006554624531418085, + 0.0065270839259028435, + 0.014902983792126179, + 0.013494668528437614, + 0.011516696773469448, + 0.03066522628068924, + 0.01235780119895935, + 0.012760031037032604, + 0.014323923736810684, + 0.013367864303290844, + 0.018348421901464462, + 0.01894899271428585, + 0.010621066205203533, + 0.01614035852253437, + 0.01726997271180153, + 0.0087217902764678, + 0.009591815061867237, + 0.009159895591437817, + 0.03157476708292961, + 0.029945537447929382, + 0.006308367010205984, + 0.009887072257697582, + 0.014429858885705471, + 0.008888905867934227, + 0.005984992254525423, + 0.006259422283619642, + 0.013559559360146523, + 0.01552730705589056, + 0.00879333633929491, + 0.008049148134887218, + 0.009506383910775185, + 0.010707509703934193, + 0.010794078931212425, + 0.008889955468475819, + 0.010329104028642178, + 0.011505196802318096, + 0.007736696396023035, + 0.007989772595465183, + 0.014770304784178734, + 0.005196769721806049, + 0.010099755600094795, + 0.007572104688733816, + 0.012684842571616173, + 0.011991703882813454, + 0.011934407986700535, + 0.02234889194369316, + 0.020558703690767288, + 0.007584048435091972, + 0.017747534438967705, + 0.014400268904864788, + 0.008360189385712147, + 0.008725558407604694, + 0.01275367010384798, + 0.021210448816418648, + 0.007091017439961433, + 0.0054336064495146275, + 0.0077464161440730095, + 0.008313666097819805, + 0.004417238757014275, + 0.006064972374588251, + 0.003236002055928111, + 0.0032000949140638113, + 0.003930548205971718, + 0.0022845102939754725, + 0.002258352702483535, + 0.0025076132733374834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.1603407859802246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-panning, wide-angle aerial shot, capturing the unmistakable aesthetic of Hayao Miyazaki's animated films, slowly sweeps across a vast, lush valley. The soft afternoon sun bathes everything in a warm golden hue, casting a soft glow over the rolling hills. A few large, fluffy, sharply defined cotton candy clouds drift across the sky. Amidst this magnificent natural landscape, a lone explorer, carrying a large, vintage travel bag, walks along a narrow dirt path winding through the meadow. Dressed in a durable beige canvas jacket, dark trousers, and sturdy hiking boots, the explorer's massive leather and thick canvas travel bag, overflowing with belongings, swayed slightly with his steady, deliberate strides. A metal water bottle and a rolled-up sleeping bag dangle from its side. The camera follows him, his back appears small and determined in the vast world, as he walks step by step towards the unknown fantasy world shrouded in mist and with strange outlines on the distant horizon. The whole picture is full of tranquil poetry and infinite yearning for an unknown journey, creating a peaceful, lonely and hopeful atmosphere.", + "seed": 15348, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_5653_seed_15348_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": -0.07436462119221687, + "vlm_logprob": -0.07436462119221687, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.07436462119221687, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016398655250668526 + ], + "opposite_logprob_tokens": [ + -4.2663984298706055 + ], + "logprob_reward": -0.016398655250668526, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016398655250668526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08776569366455078, + 0.026538359001278877, + 0.028237013146281242, + 0.09443914145231247, + 0.038851842284202576, + 0.2797371447086334, + 0.029856063425540924, + 0.34619441628456116, + 0.04627365246415138, + 0.022106684744358063 + ], + "credit_map": [ + 0.013663167133927345, + 0.012926697731018066, + 0.010259272530674934, + 0.013114546425640583, + 0.006945302244275808, + 0.010535555891692638, + 0.01157843042165041, + 0.008742709644138813, + 0.0036440156400203705, + 0.0029451351147145033, + 0.0027539001312106848, + 0.0035769189707934856, + 0.0032151956111192703, + 0.0032938318327069283, + 0.003592441324144602, + 0.0035169157199561596, + 0.0033747947309166193, + 0.004324531648308039, + 0.002234335755929351, + 0.002651557791978121, + 0.004619692452251911, + 0.0020712155383080244, + 0.006178094074130058, + 0.0027827892918139696, + 0.014868445694446564, + 0.015869276598095894, + 0.005992114078253508, + 0.007064082194119692, + 0.0200799573212862, + 0.0126477787271142, + 0.01095188781619072, + 0.006965585518628359, + 0.0075138406828045845, + 0.0035856864415109158, + 0.005169331096112728, + 0.003267245600000024, + 0.003501271130517125, + 0.00805679801851511, + 0.0031303453724831343, + 0.0046273209154605865, + 0.10911306738853455, + 0.008986837230622768, + 0.008334389887750149, + 0.004762937780469656, + 0.009163472801446915, + 0.12386882305145264, + 0.005968455225229263, + 0.009539157152175903, + 0.005856945179402828, + 0.0036309529095888138, + 0.0031389398500323296, + 0.0023950079921633005, + 0.0026119560934603214, + 0.006697359494864941, + 0.0022211261093616486, + 0.0033037748653441668, + 0.1158328428864479, + 0.015993529930710793, + 0.014726009219884872, + 0.016772128641605377, + 0.02026277221739292, + 0.1340784877538681, + 0.012173308990895748, + 0.016355307772755623, + 0.008634393103420734, + 0.002713663736358285, + 0.005624073091894388, + 0.002913517877459526, + 0.005250719375908375, + 0.009711402468383312, + 0.003317427821457386, + 0.008108450099825859, + 0.004804758355021477, + 0.001132003148086369, + 0.0019320318242534995, + 0.0012896428816020489, + 0.0028345983009785414, + 0.005526312626898289, + 0.0014235918642953038, + 0.00316374353133142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.016398655250668526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638768292963505 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007638768292963505, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638768292963505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08001723885536194, + 0.02680402621626854, + 0.019576754420995712, + 0.01790449395775795, + 0.024572480469942093, + 0.5819509625434875, + 0.13476213812828064, + 0.046843916177749634, + 0.04403068870306015, + 0.02353724278509617 + ], + "credit_map": [ + 0.012209723703563213, + 0.011661943979561329, + 0.00953423511236906, + 0.007899140939116478, + 0.008284639567136765, + 0.012163465842604637, + 0.009935194626450539, + 0.008328893221914768, + 0.0039885337464511395, + 0.003566724481061101, + 0.0027627446688711643, + 0.003473954973742366, + 0.0041750166565179825, + 0.0026101991534233093, + 0.0039982725866138935, + 0.0022285780869424343, + 0.0035274354740977287, + 0.0024555977433919907, + 0.0020606149919331074, + 0.0026565107982605696, + 0.00175562989898026, + 0.002172357402741909, + 0.002526569413021207, + 0.0024220398627221584, + 0.003172884462401271, + 0.0030695463065057993, + 0.0019221706315875053, + 0.0013036797754466534, + 0.002801857888698578, + 0.0018884273013100028, + 0.001960000256076455, + 0.0017859269864857197, + 0.005443246569484472, + 0.0026543806307017803, + 0.0018841512501239777, + 0.002017626538872719, + 0.0021500398870557547, + 0.005604115314781666, + 0.0026209750212728977, + 0.0021979466546326876, + 0.23209170997142792, + 0.020107438787817955, + 0.013330386951565742, + 0.00841379165649414, + 0.017354901880025864, + 0.2654901444911957, + 0.009400224313139915, + 0.015762407332658768, + 0.04450897499918938, + 0.007558209355920553, + 0.005202427506446838, + 0.00752284424379468, + 0.007397265639156103, + 0.05037368834018707, + 0.0054208640940487385, + 0.006777866743505001, + 0.015919959172606468, + 0.002235077554360032, + 0.0017675281269475818, + 0.002322704764083028, + 0.0024753501638770103, + 0.018692931160330772, + 0.0014518515672534704, + 0.001978509593755007, + 0.008921638131141663, + 0.0036200082395225763, + 0.003373329993337393, + 0.005751133896410465, + 0.004771322477608919, + 0.009689231403172016, + 0.002942061750218272, + 0.004961964208632708, + 0.003971832804381847, + 0.0016122196102514863, + 0.0024117205757647753, + 0.0024180670734494925, + 0.002602811437100172, + 0.004982117097824812, + 0.001807528780773282, + 0.0037309464532881975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.007638768292963505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072265915572643 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072265915572643, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072265915572643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04836798459291458, + 0.04420727118849754, + 0.01995539851486683, + 0.07906803488731384, + 0.06742142140865326, + 0.3927021026611328, + 0.09796906262636185, + 0.08537972718477249, + 0.09819331020116806, + 0.0667356327176094 + ], + "credit_map": [ + 0.006071372423321009, + 0.006743576377630234, + 0.00438869372010231, + 0.00488865002989769, + 0.005944002419710159, + 0.006121601909399033, + 0.008138809353113174, + 0.006071282550692558, + 0.00446149380877614, + 0.005481834523379803, + 0.0029972256161272526, + 0.0033902875147759914, + 0.011163443326950073, + 0.003649045480415225, + 0.008731608279049397, + 0.0043323347344994545, + 0.002869606716558337, + 0.00342083559371531, + 0.001370954792946577, + 0.001699478947557509, + 0.003123028203845024, + 0.0022336093243211508, + 0.0030216362792998552, + 0.0022162501700222492, + 0.013847497291862965, + 0.01382108312100172, + 0.005498930811882019, + 0.004915884230285883, + 0.01702873222529888, + 0.008153531700372696, + 0.009526231326162815, + 0.006276158150285482, + 0.017174694687128067, + 0.00673128804191947, + 0.006597244180738926, + 0.004035621881484985, + 0.0049463436007499695, + 0.017390785738825798, + 0.004144147504121065, + 0.006401302292943001, + 0.14777177572250366, + 0.01757817529141903, + 0.012497619725763798, + 0.009206626564264297, + 0.014849661849439144, + 0.16909120976924896, + 0.008706603199243546, + 0.013000451028347015, + 0.02242342010140419, + 0.007602018304169178, + 0.008241009898483753, + 0.011896101757884026, + 0.005416510161012411, + 0.02612980268895626, + 0.007243966683745384, + 0.009016246534883976, + 0.016617372632026672, + 0.011516869068145752, + 0.005146755371242762, + 0.012157090939581394, + 0.006728578824549913, + 0.020275890827178955, + 0.006219853647053242, + 0.006717326585203409, + 0.01925899274647236, + 0.012493805028498173, + 0.007821041159331799, + 0.013984744437038898, + 0.008083183318376541, + 0.019945954903960228, + 0.008122280240058899, + 0.008483321405947208, + 0.01983228698372841, + 0.0037479305174201727, + 0.003736875718459487, + 0.003747881157323718, + 0.006243616342544556, + 0.02173706889152527, + 0.00262068840675056, + 0.00506929587572813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.011072265915572643 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287607356905937 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287607356905937, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287607356905937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11385155469179153, + 0.06945420056581497, + 0.029063118621706963, + 0.07643923908472061, + 0.031586918979883194, + 0.2623416483402252, + 0.09992202371358871, + 0.20093919336795807, + 0.020379360765218735, + 0.09602276980876923 + ], + "credit_map": [ + 0.016920147463679314, + 0.016160450875759125, + 0.009290248155593872, + 0.013721361756324768, + 0.012768450193107128, + 0.013522567227482796, + 0.019719315692782402, + 0.011749007739126682, + 0.011914577335119247, + 0.012892473489046097, + 0.004836059175431728, + 0.004679023288190365, + 0.008148455061018467, + 0.005945724900811911, + 0.014480807818472385, + 0.006557071581482887, + 0.003540194593369961, + 0.003438228741288185, + 0.0025313447695225477, + 0.004145316779613495, + 0.0030873550567775965, + 0.0033937704283744097, + 0.005400241352617741, + 0.0035266675986349583, + 0.015539078041911125, + 0.016905510798096657, + 0.005984187591820955, + 0.004273940809071064, + 0.010330098681151867, + 0.009704737924039364, + 0.007006566505879164, + 0.006695117801427841, + 0.006782814394682646, + 0.0034184660762548447, + 0.0030958966817706823, + 0.0025812885724008083, + 0.0029641585424542427, + 0.0065324995666742325, + 0.0026749586686491966, + 0.0035368355456739664, + 0.09813841432332993, + 0.012161897495388985, + 0.007807958871126175, + 0.007079828064888716, + 0.009682773612439632, + 0.11219505220651627, + 0.006796273402869701, + 0.008479450829327106, + 0.01463955920189619, + 0.01772857829928398, + 0.007827702909708023, + 0.015138470567762852, + 0.009929724037647247, + 0.011576700024306774, + 0.01387843582779169, + 0.009202864952385426, + 0.048128530383110046, + 0.025118159130215645, + 0.010043997317552567, + 0.022113004699349403, + 0.014231841079890728, + 0.0559181347489357, + 0.012550514191389084, + 0.012835018336772919, + 0.0025025149807333946, + 0.003167870454490185, + 0.0022788336500525475, + 0.003890200750902295, + 0.001905746408738196, + 0.0022474932484328747, + 0.002205249620601535, + 0.002181452000513673, + 0.022309591993689537, + 0.004755661357194185, + 0.007121017202734947, + 0.005166041199117899, + 0.011952690780162811, + 0.02690962888300419, + 0.005721725057810545, + 0.012086411938071251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.023287607356905937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342580914497375 + ], + "opposite_logprob_tokens": [ + -1.3134257793426514 + ], + "logprob_reward": -0.31342580914497375, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342580914497375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11562871932983398, + 0.0755021870136261, + 0.07667316496372223, + 0.10683845728635788, + 0.027498818933963776, + 0.22987963259220123, + 0.13325099647045135, + 0.14855638146400452, + 0.030272159725427628, + 0.05589945614337921 + ], + "credit_map": [ + 0.016060523688793182, + 0.016097908839583397, + 0.0068399724550545216, + 0.01158849522471428, + 0.017137307673692703, + 0.008407123386859894, + 0.02819962054491043, + 0.011297762393951416, + 0.010192000307142735, + 0.008383763954043388, + 0.007217560429126024, + 0.010654853656888008, + 0.013521275483071804, + 0.00621039280667901, + 0.012786033563315868, + 0.006536305882036686, + 0.012185752391815186, + 0.013097383081912994, + 0.005024491343647242, + 0.004412252921611071, + 0.012766852974891663, + 0.005733572877943516, + 0.017048509791493416, + 0.006404351443052292, + 0.014568032696843147, + 0.01351107470691204, + 0.007159696891903877, + 0.00803348608314991, + 0.02825932204723358, + 0.010198756121098995, + 0.01819753088057041, + 0.006910557392984629, + 0.004728848580271006, + 0.002314003650099039, + 0.003111128695309162, + 0.0024196498561650515, + 0.0031444851774722338, + 0.0054012262262403965, + 0.0028084348887205124, + 0.0035710404627025127, + 0.08333118259906769, + 0.011200887151062489, + 0.006785233039408922, + 0.009362964890897274, + 0.007790745701640844, + 0.09400170296430588, + 0.007784874644130468, + 0.009622047655284405, + 0.042628213763237, + 0.010735461488366127, + 0.006068538408726454, + 0.008631275966763496, + 0.006258716341108084, + 0.04722186550498009, + 0.0057517956010997295, + 0.0059551214799284935, + 0.04871455207467079, + 0.008570482954382896, + 0.005409660283476114, + 0.008954780176281929, + 0.00848118681460619, + 0.056899331510066986, + 0.005298381205648184, + 0.006228017620742321, + 0.006519435439258814, + 0.0033569769002497196, + 0.002313600154593587, + 0.0043588969856500626, + 0.0022219503298401833, + 0.0064421421848237514, + 0.002325081266462803, + 0.0027340759988874197, + 0.014503749087452888, + 0.00643925042822957, + 0.0026505477726459503, + 0.005969619378447533, + 0.004846634808927774, + 0.014759212732315063, + 0.003171777818351984, + 0.0035586650483310223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.31342580914497375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-panning, wide-angle aerial shot, capturing the unmistakable aesthetic of Hayao Miyazaki's animated films, slowly sweeps across a vast, lush valley. The soft afternoon sun bathes everything in a warm golden hue, casting a soft glow over the rolling hills. A few large, fluffy, sharply defined cotton candy clouds drift across the sky. Amidst this magnificent natural landscape, a lone explorer, carrying a large, vintage travel bag, walks along a narrow dirt path winding through the meadow. Dressed in a durable beige canvas jacket, dark trousers, and sturdy hiking boots, the explorer's massive leather and thick canvas travel bag, overflowing with belongings, swayed slightly with his steady, deliberate strides. A metal water bottle and a rolled-up sleeping bag dangle from its side. The camera follows him, his back appears small and determined in the vast world, as he walks step by step towards the unknown fantasy world shrouded in mist and with strange outlines on the distant horizon. The whole picture is full of tranquil poetry and infinite yearning for an unknown journey, creating a peaceful, lonely and hopeful atmosphere.", + "seed": 15349, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_6/index_5653_seed_15349_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": -0.0841233167797327, + "vlm_logprob": -0.0841233167797327, + "vlm_yes_no_margin": 3.450000011920929, + "vlm_token_credit": -0.0841233167797327, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013902021571993828 + ], + "opposite_logprob_tokens": [ + -4.513902187347412 + ], + "logprob_reward": -0.013902021571993828, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013902021571993828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3046359121799469, + 0.07877103984355927, + 0.16297559440135956, + 0.04542072117328644, + 0.1612280011177063, + 0.1123649850487709, + 0.03592026233673096, + 0.03992180898785591, + 0.03162897005677223, + 0.027132678776979446 + ], + "credit_map": [ + 0.04130427539348602, + 0.03032226487994194, + 0.05584225803613663, + 0.02938258834183216, + 0.025496220216155052, + 0.04957928881049156, + 0.03821079432964325, + 0.03449823707342148, + 0.010703321546316147, + 0.009273042902350426, + 0.007223568391054869, + 0.010180790908634663, + 0.008573333732783794, + 0.012323015369474888, + 0.011879227124154568, + 0.008614744059741497, + 0.0197136327624321, + 0.01183487381786108, + 0.03680074214935303, + 0.02492673508822918, + 0.010654974728822708, + 0.020633766427636147, + 0.016990795731544495, + 0.021420065313577652, + 0.005136218387633562, + 0.0046982718631625175, + 0.008174847811460495, + 0.005681066773831844, + 0.0031852691899985075, + 0.006058871746063232, + 0.006456669885665178, + 0.006029506679624319, + 0.01456972025334835, + 0.020491523668169975, + 0.0179889053106308, + 0.047719113528728485, + 0.00808085035532713, + 0.012610051780939102, + 0.018599817529320717, + 0.021168021485209465, + 0.012632269412279129, + 0.010980339720845222, + 0.02672668732702732, + 0.010298078879714012, + 0.00799404177814722, + 0.016154244542121887, + 0.011493472382426262, + 0.016085851937532425, + 0.004459867719560862, + 0.003295064205303788, + 0.003533350769430399, + 0.007717222906649113, + 0.0032084835693240166, + 0.005825941916555166, + 0.00377531535923481, + 0.004105016123503447, + 0.0049326675944030285, + 0.004907037131488323, + 0.0037404620088636875, + 0.0076286219991743565, + 0.0038293874822556973, + 0.004618498496711254, + 0.00574610335752368, + 0.004519027657806873, + 0.005057351663708687, + 0.0033187095541507006, + 0.003936664201319218, + 0.0034135933965444565, + 0.0025457406882196665, + 0.007316244766116142, + 0.0031073251739144325, + 0.0029333431739360094, + 0.00440199812874198, + 0.002939643571153283, + 0.0032486184500157833, + 0.003243843326345086, + 0.0019763209857046604, + 0.004656412173062563, + 0.003298212541267276, + 0.003367630997672677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.013902021571993828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39131954312324524, + 0.07667288929224014, + 0.1514880508184433, + 0.08388912677764893, + 0.09555257856845856, + 0.04783501476049423, + 0.04569636657834053, + 0.05260469764471054, + 0.023264575749635696, + 0.03167711943387985 + ], + "credit_map": [ + 0.0765148401260376, + 0.03688882663846016, + 0.06772259622812271, + 0.028770197182893753, + 0.03327291086316109, + 0.07397086918354034, + 0.03214268013834953, + 0.04203663393855095, + 0.011304936371743679, + 0.010097818449139595, + 0.007276119198650122, + 0.010284699499607086, + 0.008913051337003708, + 0.011623396538197994, + 0.009435445070266724, + 0.007737420964986086, + 0.030575359240174294, + 0.011899787001311779, + 0.02382412739098072, + 0.02205238863825798, + 0.008615904487669468, + 0.024540387094020844, + 0.011487133800983429, + 0.018492965027689934, + 0.012811537832021713, + 0.006122058257460594, + 0.012467731721699238, + 0.008854449726641178, + 0.006467226427048445, + 0.020434899255633354, + 0.008010083809494972, + 0.008721129968762398, + 0.017235882580280304, + 0.008700425736606121, + 0.008925757370889187, + 0.0106048583984375, + 0.007275530137121677, + 0.02848539687693119, + 0.006288766395300627, + 0.008035950362682343, + 0.006551503203809261, + 0.00486814184114337, + 0.009824400767683983, + 0.0038002978544682264, + 0.00359754404053092, + 0.006847345270216465, + 0.005206682253628969, + 0.007139093242585659, + 0.006890124641358852, + 0.0039567188359797, + 0.004489042796194553, + 0.006388610694557428, + 0.00432555191218853, + 0.010981231927871704, + 0.003557547926902771, + 0.005107537843286991, + 0.00717891426756978, + 0.0050340392626821995, + 0.007408561185002327, + 0.005385896191000938, + 0.00513842748478055, + 0.009748246520757675, + 0.006721863988786936, + 0.0059887426905334, + 0.0035878843627870083, + 0.002522859023883939, + 0.0027211420238018036, + 0.0018695619655773044, + 0.002283445792272687, + 0.0055723609402775764, + 0.002081057522445917, + 0.002626261208206415, + 0.004019757267087698, + 0.0025867761578410864, + 0.00539850490167737, + 0.0035856193862855434, + 0.0021834406070411205, + 0.004229681566357613, + 0.00483038742095232, + 0.00484295142814517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.020592818036675453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192244037985802 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014192244037985802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192244037985802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36700600385665894, + 0.14431564509868622, + 0.11152136325836182, + 0.060375675559043884, + 0.12541162967681885, + 0.039180513471364975, + 0.0429718904197216, + 0.05242875963449478, + 0.029429646208882332, + 0.027358822524547577 + ], + "credit_map": [ + 0.043698813766241074, + 0.04957174137234688, + 0.053001098334789276, + 0.04526422172784805, + 0.03518546000123024, + 0.052761685103178024, + 0.0453394316136837, + 0.04218357428908348, + 0.022790122777223587, + 0.02157752215862274, + 0.013838418759405613, + 0.012211221270263195, + 0.019777903333306313, + 0.022885899990797043, + 0.017634661868214607, + 0.01359990332275629, + 0.013742413371801376, + 0.017304085195064545, + 0.01632579416036606, + 0.00996023416519165, + 0.012840300798416138, + 0.012253182008862495, + 0.01711978390812874, + 0.011975574307143688, + 0.007882297039031982, + 0.006590475793927908, + 0.007138343527913094, + 0.00937645509839058, + 0.006731873378157616, + 0.008826196193695068, + 0.008015213534235954, + 0.005814820062369108, + 0.018691647797822952, + 0.011426028795540333, + 0.015578476712107658, + 0.017768198624253273, + 0.009747768752276897, + 0.01670422963798046, + 0.019609730690717697, + 0.01588555797934532, + 0.004719628486782312, + 0.0051063974387943745, + 0.004623265936970711, + 0.0047828336246311665, + 0.0039023293647915125, + 0.0040617226622998714, + 0.007817355915904045, + 0.0041669816710054874, + 0.005004303529858589, + 0.005065362900495529, + 0.004397825337946415, + 0.007129225879907608, + 0.00337173230946064, + 0.006479877047240734, + 0.006265603005886078, + 0.005257964134216309, + 0.0062382216565310955, + 0.007779544219374657, + 0.005420510657131672, + 0.006691978313028812, + 0.005389122292399406, + 0.006131866946816444, + 0.008460763841867447, + 0.006316752173006535, + 0.003651826409623027, + 0.00395785178989172, + 0.003111062105745077, + 0.0043742540292441845, + 0.00307140639051795, + 0.003826896660029888, + 0.004282322246581316, + 0.003154025413095951, + 0.003034920897334814, + 0.002603733679279685, + 0.00539707625284791, + 0.0030837692320346832, + 0.001993143931031227, + 0.003008024301379919, + 0.003840240417048335, + 0.004397913813591003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014192244037985802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294363170862198 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.023294363170862198, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294363170862198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22428062558174133, + 0.11184034496545792, + 0.09613869339227676, + 0.18709619343280792, + 0.13906218111515045, + 0.04455496743321419, + 0.03445045277476311, + 0.05421849712729454, + 0.04748787358403206, + 0.06087014451622963 + ], + "credit_map": [ + 0.02559269592165947, + 0.03192362189292908, + 0.03782712668180466, + 0.025157218798995018, + 0.024280782788991928, + 0.020959537476301193, + 0.02924310974776745, + 0.029296552762389183, + 0.015614542178809643, + 0.017383906990289688, + 0.01346045732498169, + 0.011407604441046715, + 0.012772062793374062, + 0.014976288191974163, + 0.014497151598334312, + 0.011728338897228241, + 0.014047060161828995, + 0.012915404513478279, + 0.010931236669421196, + 0.008297878317534924, + 0.010436381213366985, + 0.014772221446037292, + 0.014511331915855408, + 0.01022719033062458, + 0.02168995514512062, + 0.021682552993297577, + 0.02107309177517891, + 0.042073577642440796, + 0.016892556101083755, + 0.016197944059967995, + 0.026856161653995514, + 0.02063036523759365, + 0.018296102061867714, + 0.027335558086633682, + 0.015561581589281559, + 0.022200170904397964, + 0.011006567627191544, + 0.01864110864698887, + 0.013356011360883713, + 0.012665082700550556, + 0.00688640121370554, + 0.008290095254778862, + 0.00483362004160881, + 0.005437413230538368, + 0.0042653800919651985, + 0.005463590379804373, + 0.004857433959841728, + 0.0045210374519228935, + 0.004730759654194117, + 0.004282863810658455, + 0.004183489829301834, + 0.003295859321951866, + 0.0035838568583130836, + 0.006543642375618219, + 0.004105711355805397, + 0.0037242749240249395, + 0.006896189413964748, + 0.006896767299622297, + 0.006602805107831955, + 0.005423331633210182, + 0.00486975722014904, + 0.010758684016764164, + 0.0068101040087640285, + 0.00596086448058486, + 0.007300708442926407, + 0.0049094147980213165, + 0.004500427283346653, + 0.005616549402475357, + 0.004117388278245926, + 0.011010843329131603, + 0.005701048299670219, + 0.004331502597779036, + 0.01024964265525341, + 0.00403591338545084, + 0.00849524699151516, + 0.005417528562247753, + 0.00317038013599813, + 0.014092210680246353, + 0.006781587842851877, + 0.008627638220787048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.023294363170862198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 wide-angle aerial view of a vast, lush valley with rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a soft afternoon sun casting a warm golden hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a lone explorer wearing a beige canvas jacket, dark trousers, and sturdy hiking boots walking on a narrow dirt path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the explorer carry a large vintage travel bag with a metal water bottle and a rolled-up sleeping bag dangling from its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow right-panning movement following the explorer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34863513708114624 + ], + "opposite_logprob_tokens": [ + -1.223635196685791 + ], + "logprob_reward": -0.34863513708114624, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34863513708114624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19300058484077454, + 0.13893063366413116, + 0.19767582416534424, + 0.07021921128034592, + 0.13518397510051727, + 0.03232460469007492, + 0.056627027690410614, + 0.034007444977760315, + 0.036303143948316574, + 0.10572751611471176 + ], + "credit_map": [ + 0.02928013913333416, + 0.019092300906777382, + 0.02077702432870865, + 0.018403753638267517, + 0.015252294018864632, + 0.050349168479442596, + 0.023220833390951157, + 0.01662508398294449, + 0.019522476941347122, + 0.021515093743801117, + 0.010159400291740894, + 0.01425386406481266, + 0.02043684385716915, + 0.019253747537732124, + 0.020331522449851036, + 0.013457686640322208, + 0.026452133432030678, + 0.017327237874269485, + 0.04219336807727814, + 0.021916760131716728, + 0.016766386106610298, + 0.027685724198818207, + 0.020687570795416832, + 0.024646656587719917, + 0.0077656907960772514, + 0.008626185357570648, + 0.008353601209819317, + 0.01195153035223484, + 0.006280452944338322, + 0.011650843545794487, + 0.008453543297946453, + 0.00713736517354846, + 0.021462056785821915, + 0.01959730125963688, + 0.014343584887683392, + 0.028223976492881775, + 0.011413964442908764, + 0.014724013395607471, + 0.01280198898166418, + 0.012617094442248344, + 0.0039005570579320192, + 0.005021661054342985, + 0.005734246224164963, + 0.0028167252894490957, + 0.0027006256859749556, + 0.0054650502279400826, + 0.0031432481482625008, + 0.00354249426163733, + 0.007293637841939926, + 0.007878946140408516, + 0.004936791956424713, + 0.010283678770065308, + 0.00657983822748065, + 0.008768212050199509, + 0.005143248476088047, + 0.005742679350078106, + 0.003931842278689146, + 0.003150555305182934, + 0.006551533471792936, + 0.005353335291147232, + 0.0021377291996032, + 0.0041180080734193325, + 0.004018788225948811, + 0.004745655693113804, + 0.004733704030513763, + 0.0044899024069309235, + 0.0027532584499567747, + 0.0049545117653906345, + 0.00394481373950839, + 0.006192599423229694, + 0.005059991963207722, + 0.004174365196377039, + 0.014427403919398785, + 0.01290903426706791, + 0.011675243265926838, + 0.014928048476576805, + 0.010204854421317577, + 0.011939938180148602, + 0.01724446378648281, + 0.012398538179695606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.34863513708114624 + } + ], + "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_7/index_tensor([6200, 6200])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_tensor([6200, 6200])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5cf7921ad240c823d5a4dee685afda1ec6102a47 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_tensor([6200, 6200])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "terrifying first-person perspective shot, shot from the damp, filthy floor of a forgotten sewer, rises from an extreme low-angle perspective in an utterly terrifying manner. The camera remains still, emphasizing a sense of entrapment and helplessness. The surroundings are suffocatingly dark, their curved, leaking concrete walls coated in algae and grime in a chilling color temperature. A harsh, dusty light, streaming down from a high grate, pierces the darkness, illuminating the scene's horrific centerpiece: a grotesque, living chandelier composed of countless sewer rats. Their damp, shaggy fur—a sickening gray and brown—clings to their writhing bodies. Their long, naked tails are horribly knotted at the tip, forming a single, dense, fleshy connection from which the entire writhing mass hangs. The abominations perform a slow, steady, hypnotic rotation on a fixed axis, performing a macabre dance within the solitary shaft of light. As the crowd turned, individual rats squirmed and struggled futilely, their microscopic movements sending constant, sickening ripples through the entire mass. The harsh light caught the damp sheen of their fur, and most terrifying of all, a thousand tiny, black, beady eyes gleamed with malevolence as they stared into the darkness, creating an atmosphere of pure, oppressive fear and visceral loathing.", + "seed": 4437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4437_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": -0.06824567478615791, + "vlm_logprob": -0.06824567478615791, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.06824567478615791, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14179232716560364 + ], + "opposite_logprob_tokens": [ + -2.2667922973632812 + ], + "logprob_reward": -0.14179232716560364, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.14179232716560364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38110336661338806, + 0.08701678365468979, + 0.07767634838819504, + 0.04218435287475586, + 0.08881401270627975, + 0.09464766085147858, + 0.11422646045684814, + 0.03041057661175728, + 0.03897843137383461, + 0.0449419766664505 + ], + "credit_map": [ + 0.06160295754671097, + 0.034497011452913284, + 0.047354806214571, + 0.04259967803955078, + 0.04694337025284767, + 0.07208537310361862, + 0.04587540403008461, + 0.03014480136334896, + 0.008010162971913815, + 0.00902215950191021, + 0.02186465635895729, + 0.01364852488040924, + 0.010227186605334282, + 0.008247909136116505, + 0.007260013371706009, + 0.008736178278923035, + 0.006601396482437849, + 0.0072649577632546425, + 0.024924542754888535, + 0.010587085038423538, + 0.005869231186807156, + 0.010614677332341671, + 0.006160909775644541, + 0.005653548054397106, + 0.004729748237878084, + 0.0038302482571452856, + 0.008020161651074886, + 0.005683089606463909, + 0.004133334383368492, + 0.0053580463863909245, + 0.004537214059382677, + 0.005892516579478979, + 0.01237461157143116, + 0.0038264496251940727, + 0.00818277895450592, + 0.008365022018551826, + 0.0045468476600945, + 0.010355676524341106, + 0.0073682283982634544, + 0.03379440680146217, + 0.021004067733883858, + 0.006613365840166807, + 0.015425087884068489, + 0.01819203980267048, + 0.004486669786274433, + 0.010102365165948868, + 0.00688821729272604, + 0.011935855261981487, + 0.015188643708825111, + 0.006926337722688913, + 0.017664918676018715, + 0.015389797277748585, + 0.008148742839694023, + 0.018323998898267746, + 0.009200136177241802, + 0.02338389679789543, + 0.0028550124261528254, + 0.0026355329900979996, + 0.0031300042755901814, + 0.0038164216093719006, + 0.0032418416813015938, + 0.0037632444873452187, + 0.0037901196628808975, + 0.007178403437137604, + 0.004537888336926699, + 0.0025876990985125303, + 0.005099561531096697, + 0.004035120829939842, + 0.0034873189870268106, + 0.004816771950572729, + 0.0036313864402472973, + 0.010782686062157154, + 0.008454854600131512, + 0.0025504250079393387, + 0.007101323455572128, + 0.003726048395037651, + 0.0027489238418638706, + 0.012123452499508858, + 0.003349738195538521, + 0.004887216258794069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.14179232716560364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422479934990406 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.01422479934990406, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422479934990406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24052107334136963, + 0.062093641608953476, + 0.05635654553771019, + 0.05693709850311279, + 0.0988956093788147, + 0.11698281764984131, + 0.21947647631168365, + 0.034595832228660583, + 0.03423785790801048, + 0.07990299165248871 + ], + "credit_map": [ + 0.03252415731549263, + 0.02437461167573929, + 0.028292426839470863, + 0.021668070927262306, + 0.02765137143433094, + 0.05079016461968422, + 0.027822628617286682, + 0.02739766612648964, + 0.0059052882716059685, + 0.005718280095607042, + 0.018398452550172806, + 0.00953450333327055, + 0.005918870214372873, + 0.0056315828114748, + 0.004417191259562969, + 0.006569478660821915, + 0.004685028921812773, + 0.005053303204476833, + 0.014240999706089497, + 0.007198111619800329, + 0.0047219437547028065, + 0.007454290520399809, + 0.005717422813177109, + 0.007285448256880045, + 0.007952114567160606, + 0.003138059750199318, + 0.013539479114115238, + 0.006037399638444185, + 0.0035697398707270622, + 0.00824593473225832, + 0.005146971438080072, + 0.009307404980063438, + 0.013734925538301468, + 0.004887043964117765, + 0.005085495766252279, + 0.008991419337689877, + 0.003323639277368784, + 0.013527296483516693, + 0.0089491568505764, + 0.040396638214588165, + 0.026161110028624535, + 0.008214541710913181, + 0.019860917702317238, + 0.016765384003520012, + 0.006374348886311054, + 0.012531401589512825, + 0.008822062984108925, + 0.018253052607178688, + 0.02319442667067051, + 0.014089738950133324, + 0.03669455274939537, + 0.02999231405556202, + 0.016980005428195, + 0.04446573555469513, + 0.026258554309606552, + 0.027801156044006348, + 0.0036984176840633154, + 0.003102969378232956, + 0.005601378157734871, + 0.004487546626478434, + 0.006178637966513634, + 0.00340547738596797, + 0.003071103012189269, + 0.005050304811447859, + 0.002504907315596938, + 0.002759837545454502, + 0.004480786621570587, + 0.0047676037065684795, + 0.0036944583989679813, + 0.006012229714542627, + 0.0053069107234478, + 0.004711130168288946, + 0.0166366808116436, + 0.003891794476658106, + 0.005776699632406235, + 0.006437711417675018, + 0.0045199706219136715, + 0.026420501992106438, + 0.007849641144275665, + 0.008369995281100273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.01422479934990406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039450466632843 + ], + "opposite_logprob_tokens": [ + -1.910394549369812 + ], + "logprob_reward": -0.16039450466632843, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039450466632843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26706838607788086, + 0.07892579585313797, + 0.08602150529623032, + 0.05409169942140579, + 0.040935683995485306, + 0.16603712737560272, + 0.15331649780273438, + 0.04637041315436363, + 0.04040350392460823, + 0.0668293684720993 + ], + "credit_map": [ + 0.03390801325440407, + 0.029300615191459656, + 0.03588295355439186, + 0.04027193412184715, + 0.03887277841567993, + 0.02954671159386635, + 0.03358611837029457, + 0.025699254125356674, + 0.008689594455063343, + 0.006559687666594982, + 0.01922859251499176, + 0.011605554260313511, + 0.010126844979822636, + 0.008845706470310688, + 0.00706864520907402, + 0.006801167502999306, + 0.006737858988344669, + 0.005500724073499441, + 0.029742645099759102, + 0.011904465965926647, + 0.006592721678316593, + 0.012600287795066833, + 0.005697494838386774, + 0.007245312444865704, + 0.004959688987582922, + 0.004968698602169752, + 0.009113299660384655, + 0.0062127746641635895, + 0.0035697186831384897, + 0.008725835010409355, + 0.006628484930843115, + 0.00991320051252842, + 0.00458341371268034, + 0.002692724345251918, + 0.004941519349813461, + 0.00395742803812027, + 0.004734717775136232, + 0.004925285466015339, + 0.0033649192191660404, + 0.011735675856471062, + 0.03430096060037613, + 0.007041314151138067, + 0.014867079444229603, + 0.023540759459137917, + 0.00550791947171092, + 0.012518497183918953, + 0.008589809760451317, + 0.05967078357934952, + 0.016093840822577477, + 0.008966510184109211, + 0.036737412214279175, + 0.021384606137871742, + 0.010760404169559479, + 0.025385575369000435, + 0.01343529112637043, + 0.020552853122353554, + 0.004224450793117285, + 0.005055909510701895, + 0.006478165276348591, + 0.0061636799946427345, + 0.00436011329293251, + 0.006753211375325918, + 0.00552879273891449, + 0.007806089241057634, + 0.004254448693245649, + 0.0028348977211862803, + 0.003430460812523961, + 0.005309839267283678, + 0.004864532966166735, + 0.006376473233103752, + 0.006338140927255154, + 0.0069947089068591595, + 0.013996158726513386, + 0.003967728931456804, + 0.004722483456134796, + 0.006329293362796307, + 0.004237605258822441, + 0.020210379734635353, + 0.008279239758849144, + 0.005086476448923349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.16039450466632843 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015190026024356484 + ], + "opposite_logprob_tokens": [ + -6.501519203186035 + ], + "logprob_reward": -0.0015190026024356484, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015190026024356484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19756454229354858, + 0.12959235906600952, + 0.16876035928726196, + 0.05520638823509216, + 0.03473389148712158, + 0.12697601318359375, + 0.11177187412977219, + 0.02855643257498741, + 0.0427214689552784, + 0.10411665588617325 + ], + "credit_map": [ + 0.028092719614505768, + 0.018863961100578308, + 0.03088444471359253, + 0.028168629854917526, + 0.02163388952612877, + 0.02546807751059532, + 0.02737489528954029, + 0.017077945172786713, + 0.009322329424321651, + 0.01549200527369976, + 0.03333402797579765, + 0.02072998695075512, + 0.01388680562376976, + 0.0106274439021945, + 0.011613565497100353, + 0.014586200937628746, + 0.013256026431918144, + 0.012391593307256699, + 0.05768432840704918, + 0.024036400020122528, + 0.010589375160634518, + 0.02656896598637104, + 0.013461505994200706, + 0.010772162117064, + 0.006600853521376848, + 0.0049822633154690266, + 0.009608299471437931, + 0.006837931461632252, + 0.0050672502256929874, + 0.006666214670985937, + 0.007832014001905918, + 0.007611564826220274, + 0.0032597051467746496, + 0.003345477394759655, + 0.006290158722549677, + 0.004160001873970032, + 0.004448520019650459, + 0.003435153281316161, + 0.004616966005414724, + 0.005177914164960384, + 0.014919869601726532, + 0.00709658395498991, + 0.011702091433107853, + 0.020553134381771088, + 0.00830699224025011, + 0.012308334931731224, + 0.009257799945771694, + 0.04283120483160019, + 0.010076293721795082, + 0.005438606720417738, + 0.020031988620758057, + 0.01640942692756653, + 0.008340798318386078, + 0.01626412197947502, + 0.011555306613445282, + 0.023655345663428307, + 0.0030065784230828285, + 0.002389877336099744, + 0.0037468739319592714, + 0.0037011413369327784, + 0.003635081695392728, + 0.0037712363991886377, + 0.0033672451972961426, + 0.004938400350511074, + 0.004202196840196848, + 0.003124963492155075, + 0.00554936658591032, + 0.005499796476215124, + 0.0064179860055446625, + 0.007954169996082783, + 0.0050163897685706615, + 0.004956603050231934, + 0.022481005638837814, + 0.0046453895047307014, + 0.0075490036979317665, + 0.0068068779073655605, + 0.004319773521274328, + 0.038534075021743774, + 0.00625841086730361, + 0.01352212019264698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0015190026024356484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297740146517754 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297740146517754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297740146517754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2659582495689392, + 0.09649897366762161, + 0.10347603261470795, + 0.040654052048921585, + 0.11929136514663696, + 0.0692158043384552, + 0.076479472219944, + 0.09958615154027939, + 0.034486379474401474, + 0.09435354173183441 + ], + "credit_map": [ + 0.031955521553754807, + 0.02274097129702568, + 0.04072271287441254, + 0.040363211184740067, + 0.03291366621851921, + 0.037694696336984634, + 0.04019869118928909, + 0.019368795678019524, + 0.009626097977161407, + 0.012457476928830147, + 0.018371835350990295, + 0.014162256382405758, + 0.013581220991909504, + 0.009301764890551567, + 0.009141278453171253, + 0.009857038967311382, + 0.009449759498238564, + 0.010246430523693562, + 0.021439572796225548, + 0.01158516388386488, + 0.015021095052361488, + 0.011891685426235199, + 0.00997268408536911, + 0.013869640417397022, + 0.003992005251348019, + 0.003817999502643943, + 0.005550008732825518, + 0.004283449146896601, + 0.0059815323911607265, + 0.006233563646674156, + 0.005638686940073967, + 0.005156802944839001, + 0.01787455379962921, + 0.0062895058654248714, + 0.012204672209918499, + 0.01255737990140915, + 0.005773541051894426, + 0.010701078921556473, + 0.01297122985124588, + 0.040919408202171326, + 0.012281030416488647, + 0.003207790432497859, + 0.004966613370925188, + 0.00912014301866293, + 0.0029816010501235723, + 0.006857180502265692, + 0.004661127924919128, + 0.025140317156910896, + 0.008946307003498077, + 0.004666473716497421, + 0.005258496385067701, + 0.011930398643016815, + 0.006061600986868143, + 0.009851699694991112, + 0.006761932745575905, + 0.023002566769719124, + 0.008818178437650204, + 0.011453655548393726, + 0.012051856145262718, + 0.012729750946164131, + 0.016036340966820717, + 0.009581780061125755, + 0.01917693391442299, + 0.009737655520439148, + 0.003384123556315899, + 0.0025716060772538185, + 0.0068817054852843285, + 0.004035152029246092, + 0.004846231546252966, + 0.0054015908390283585, + 0.004298221319913864, + 0.003067746525630355, + 0.008004027418792248, + 0.006930207367986441, + 0.018746351823210716, + 0.01046342309564352, + 0.0068105715326964855, + 0.012661824002861977, + 0.010253864340484142, + 0.020483264699578285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.023297740146517754 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "terrifying first-person perspective shot, shot from the damp, filthy floor of a forgotten sewer, rises from an extreme low-angle perspective in an utterly terrifying manner. The camera remains still, emphasizing a sense of entrapment and helplessness. The surroundings are suffocatingly dark, their curved, leaking concrete walls coated in algae and grime in a chilling color temperature. A harsh, dusty light, streaming down from a high grate, pierces the darkness, illuminating the scene's horrific centerpiece: a grotesque, living chandelier composed of countless sewer rats. Their damp, shaggy fur—a sickening gray and brown—clings to their writhing bodies. Their long, naked tails are horribly knotted at the tip, forming a single, dense, fleshy connection from which the entire writhing mass hangs. The abominations perform a slow, steady, hypnotic rotation on a fixed axis, performing a macabre dance within the solitary shaft of light. As the crowd turned, individual rats squirmed and struggled futilely, their microscopic movements sending constant, sickening ripples through the entire mass. The harsh light caught the damp sheen of their fur, and most terrifying of all, a thousand tiny, black, beady eyes gleamed with malevolence as they stared into the darkness, creating an atmosphere of pure, oppressive fear and visceral loathing.", + "seed": 4438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4438_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": -0.016927654715254904, + "vlm_logprob": -0.016927654715254904, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.016927654715254904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04582695662975311 + ], + "opposite_logprob_tokens": [ + -3.5458269119262695 + ], + "logprob_reward": -0.04582695662975311, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.04582695662975311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10614776611328125, + 0.050313614308834076, + 0.04793727770447731, + 0.047124866396188736, + 0.07191630452871323, + 0.09180907905101776, + 0.08698190748691559, + 0.4575197696685791, + 0.015489806421101093, + 0.024759646505117416 + ], + "credit_map": [ + 0.009265649132430553, + 0.013547324575483799, + 0.013273595832288265, + 0.014698240906000137, + 0.012117558158934116, + 0.016122817993164062, + 0.017642349004745483, + 0.009480226784944534, + 0.0037748925387859344, + 0.0032242697197943926, + 0.014098449610173702, + 0.0047373720444738865, + 0.004309783689677715, + 0.003549902467057109, + 0.006430340465158224, + 0.010188604705035686, + 0.004017537925392389, + 0.006361682899296284, + 0.006244116462767124, + 0.003977712243795395, + 0.013170180842280388, + 0.0047380863688886166, + 0.0047507998533546925, + 0.00467716297134757, + 0.00754630146548152, + 0.006880556233227253, + 0.0072059170342981815, + 0.005373413674533367, + 0.004343830049037933, + 0.005751687567681074, + 0.005330419633537531, + 0.004692741669714451, + 0.007597828283905983, + 0.013705535791814327, + 0.010162385180592537, + 0.009070439264178276, + 0.007900458760559559, + 0.005499444901943207, + 0.011171165853738785, + 0.006809047423303127, + 0.01218076329678297, + 0.007653937209397554, + 0.013802234083414078, + 0.008514407090842724, + 0.004194758832454681, + 0.010253516025841236, + 0.008574694395065308, + 0.026634765788912773, + 0.014372386038303375, + 0.006236283574253321, + 0.012231770902872086, + 0.008110806345939636, + 0.007862919010221958, + 0.00921870768070221, + 0.008022681809961796, + 0.020926354452967644, + 0.03707532584667206, + 0.06820172816514969, + 0.057356979697942734, + 0.08210395276546478, + 0.0833095982670784, + 0.031689029186964035, + 0.058330047875642776, + 0.03945310413837433, + 0.0026359609328210354, + 0.002092953771352768, + 0.0021933040115982294, + 0.0019952277652919292, + 0.0015024803578853607, + 0.0014445666456595063, + 0.0013479196932166815, + 0.0022773926611989737, + 0.0036997823044657707, + 0.004268784541636705, + 0.004021927714347839, + 0.0029683124739676714, + 0.0018596750451251864, + 0.002904159016907215, + 0.0022386617492884398, + 0.0027983421459794044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.04582695662975311 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565725483000278 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.012565725483000278, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565725483000278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21229682862758636, + 0.24970988929271698, + 0.05125563219189644, + 0.06309330463409424, + 0.09681957960128784, + 0.05170594900846481, + 0.05869457498192787, + 0.08822554349899292, + 0.06452915072441101, + 0.06366957724094391 + ], + "credit_map": [ + 0.01809944212436676, + 0.03125327080488205, + 0.02735733799636364, + 0.027084533125162125, + 0.025381337851285934, + 0.03659742325544357, + 0.028958996757864952, + 0.01756448857486248, + 0.016712266951799393, + 0.019153354689478874, + 0.06155041605234146, + 0.02482798881828785, + 0.022269923239946365, + 0.024599498137831688, + 0.041281864047050476, + 0.03931458294391632, + 0.0062968130223453045, + 0.008256110362708569, + 0.006868055555969477, + 0.005062624346464872, + 0.008416484110057354, + 0.004818802233785391, + 0.006375792436301708, + 0.005160949658602476, + 0.007815090008080006, + 0.010266311466693878, + 0.008787228725850582, + 0.006759097799658775, + 0.008361821062862873, + 0.006437933072447777, + 0.008599039167165756, + 0.006066787987947464, + 0.010859613306820393, + 0.010269394144415855, + 0.018530776724219322, + 0.009115739725530148, + 0.008006596006453037, + 0.010116451419889927, + 0.015144448727369308, + 0.014776550233364105, + 0.008348232135176659, + 0.007250405382364988, + 0.008798801340162754, + 0.005669092293828726, + 0.004762462805956602, + 0.004559315275400877, + 0.005311220418661833, + 0.007006416097283363, + 0.008027051575481892, + 0.005786924622952938, + 0.00873023271560669, + 0.007495678961277008, + 0.005798738915473223, + 0.004942411556839943, + 0.005594340618699789, + 0.012319196946918964, + 0.009230946190655231, + 0.013348989188671112, + 0.012757732532918453, + 0.011886658146977425, + 0.010651596821844578, + 0.008802597410976887, + 0.00994718074798584, + 0.011599849909543991, + 0.011603597551584244, + 0.009641152806580067, + 0.00976074580103159, + 0.006052340380847454, + 0.007088670507073402, + 0.00566914351657033, + 0.005375206470489502, + 0.00933829229325056, + 0.008310952223837376, + 0.0095823984593153, + 0.009574255906045437, + 0.00749021815136075, + 0.0077912043780088425, + 0.005321991164237261, + 0.008695017546415329, + 0.006903544068336487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.012565725483000278 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608525961637497 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608525961637497, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608525961637497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46085432171821594, + 0.09845836460590363, + 0.03495170176029205, + 0.0421479195356369, + 0.08797293901443481, + 0.04289732873439789, + 0.04662800952792168, + 0.07094457745552063, + 0.04467316344380379, + 0.07047168165445328 + ], + "credit_map": [ + 0.038802750408649445, + 0.041194550693035126, + 0.056104328483343124, + 0.06430089473724365, + 0.052998073399066925, + 0.10480532795190811, + 0.06104693189263344, + 0.04160146042704582, + 0.006682893726974726, + 0.009081129916012287, + 0.014452633447945118, + 0.014436944387853146, + 0.010640421882271767, + 0.01515930611640215, + 0.016469817608594894, + 0.011535207740962505, + 0.004184899386018515, + 0.003981740679591894, + 0.005183536559343338, + 0.004479190334677696, + 0.003949064761400223, + 0.004836817272007465, + 0.0036741693038493395, + 0.00466228136792779, + 0.0041158972308039665, + 0.004823006223887205, + 0.0062127127312123775, + 0.004881045315414667, + 0.006061638705432415, + 0.0033444301225245, + 0.008809161372482777, + 0.0039000280667096376, + 0.01023456547409296, + 0.0188912283629179, + 0.011891660280525684, + 0.007945852354168892, + 0.01180420070886612, + 0.00613528024405241, + 0.011460092850029469, + 0.009610051289200783, + 0.004821512382477522, + 0.0038650883361697197, + 0.006278699729591608, + 0.005358085501939058, + 0.0030692401342093945, + 0.004452576395124197, + 0.005316837225109339, + 0.009735288098454475, + 0.005830660462379456, + 0.004091273061931133, + 0.006457160227000713, + 0.0065523069351911545, + 0.006433160975575447, + 0.004330143798142672, + 0.007107434794306755, + 0.005825867876410484, + 0.006662122905254364, + 0.010451126843690872, + 0.009097326546907425, + 0.012078611180186272, + 0.01124295499175787, + 0.005910270847380161, + 0.008504564873874187, + 0.006997591815888882, + 0.007669395767152309, + 0.006244266871362925, + 0.0069867209531366825, + 0.004095252137631178, + 0.0032401150092482567, + 0.003895785426720977, + 0.0026661751326173544, + 0.00987545307725668, + 0.007834214717149734, + 0.012768973596394062, + 0.01060458179563284, + 0.009277628734707832, + 0.00813246052712202, + 0.007993597537279129, + 0.007152874954044819, + 0.006707340944558382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01608525961637497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015161458868533373 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015161458868533373, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015161458868533373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23779131472110748, + 0.08213022351264954, + 0.08688848465681076, + 0.04725305736064911, + 0.10639841109514236, + 0.07961069792509079, + 0.1444619596004486, + 0.0529622845351696, + 0.03361755236983299, + 0.12888601422309875 + ], + "credit_map": [ + 0.029133662581443787, + 0.030864525586366653, + 0.03271687030792236, + 0.026059968397021294, + 0.02743487060070038, + 0.04037756100296974, + 0.029758954420685768, + 0.021444888785481453, + 0.00596433412283659, + 0.0068612657487392426, + 0.017702743411064148, + 0.010378583334386349, + 0.008542201481759548, + 0.009167077019810677, + 0.009352426044642925, + 0.01416158676147461, + 0.008942008949816227, + 0.008388875983655453, + 0.012136092409491539, + 0.010595470666885376, + 0.017674824222922325, + 0.012372833676636219, + 0.007092603482306004, + 0.009685776196420193, + 0.005996481981128454, + 0.006563800852745771, + 0.006915392354130745, + 0.007859557867050171, + 0.004564093891531229, + 0.0037782404106110334, + 0.005751852411776781, + 0.005823633633553982, + 0.011581853963434696, + 0.021313585340976715, + 0.013959153555333614, + 0.011807025410234928, + 0.011842690408229828, + 0.008362396620213985, + 0.015471156686544418, + 0.012060548178851604, + 0.010812174528837204, + 0.013284951448440552, + 0.01109323836863041, + 0.007101752795279026, + 0.00846921093761921, + 0.010252051055431366, + 0.008286905474960804, + 0.010310414247214794, + 0.017078114673495293, + 0.012321735732257366, + 0.019661227241158485, + 0.018254615366458893, + 0.018149705603718758, + 0.014559346251189709, + 0.014584336429834366, + 0.029852885752916336, + 0.004745881073176861, + 0.00955414492636919, + 0.006635610945522785, + 0.008461774326860905, + 0.0074660032987594604, + 0.0034302708227187395, + 0.0063967518508434296, + 0.006271847058087587, + 0.00459328992292285, + 0.006265952717512846, + 0.004896895028650761, + 0.0033756482880562544, + 0.003883620025590062, + 0.0031477150041610003, + 0.0029737602453678846, + 0.0044806706719100475, + 0.01287958025932312, + 0.02427038922905922, + 0.01870833896100521, + 0.01567491888999939, + 0.01787140965461731, + 0.013261338695883751, + 0.015470068901777267, + 0.010749982669949532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0015161458868533373 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644185960292816 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644185960292816, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644185960292816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17790329456329346, + 0.11581506580114365, + 0.08966191858053207, + 0.05415472015738487, + 0.07757503539323807, + 0.11265064775943756, + 0.06565038859844208, + 0.15083833038806915, + 0.07454600930213928, + 0.0812046080827713 + ], + "credit_map": [ + 0.021292483434081078, + 0.02259051240980625, + 0.021662233397364616, + 0.0185925904661417, + 0.019922688603401184, + 0.03361395001411438, + 0.024840330705046654, + 0.015388499945402145, + 0.008737236261367798, + 0.007117884699255228, + 0.02981465496122837, + 0.01380183082073927, + 0.00789314229041338, + 0.01067535299807787, + 0.01764666847884655, + 0.020128291100263596, + 0.007412489503622055, + 0.010442547500133514, + 0.011632309295237064, + 0.00914840493351221, + 0.023050827905535698, + 0.010456093586981297, + 0.008920258842408657, + 0.008598986081779003, + 0.008703049272298813, + 0.006761117372661829, + 0.008777511306107044, + 0.006205474492162466, + 0.004682226572185755, + 0.007087002042680979, + 0.005332374945282936, + 0.0066059660166502, + 0.007659914903342724, + 0.016374386847019196, + 0.009580502286553383, + 0.009707674384117126, + 0.009209430776536465, + 0.0076455362141132355, + 0.010526349768042564, + 0.006871241144835949, + 0.014766072854399681, + 0.016038354486227036, + 0.01647152006626129, + 0.011438429355621338, + 0.01007046177983284, + 0.015160703100264072, + 0.011637784540653229, + 0.017067328095436096, + 0.008841818198561668, + 0.004106773063540459, + 0.010125808417797089, + 0.005501358769834042, + 0.005688374862074852, + 0.007284488994628191, + 0.0050941952504217625, + 0.019007569178938866, + 0.013011816889047623, + 0.03013777919113636, + 0.02028299681842327, + 0.020197736099362373, + 0.028132112696766853, + 0.011554223485291004, + 0.0159803107380867, + 0.01154135912656784, + 0.010872701182961464, + 0.012277948670089245, + 0.010440916754305363, + 0.009793638251721859, + 0.008731920272111893, + 0.006650449242442846, + 0.007585759274661541, + 0.008192676119506359, + 0.008425150066614151, + 0.013493870384991169, + 0.011705161072313786, + 0.010635275393724442, + 0.011268011294305325, + 0.008580942638218403, + 0.01024569757282734, + 0.006850502919405699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.008644185960292816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "terrifying first-person perspective shot, shot from the damp, filthy floor of a forgotten sewer, rises from an extreme low-angle perspective in an utterly terrifying manner. The camera remains still, emphasizing a sense of entrapment and helplessness. The surroundings are suffocatingly dark, their curved, leaking concrete walls coated in algae and grime in a chilling color temperature. A harsh, dusty light, streaming down from a high grate, pierces the darkness, illuminating the scene's horrific centerpiece: a grotesque, living chandelier composed of countless sewer rats. Their damp, shaggy fur—a sickening gray and brown—clings to their writhing bodies. Their long, naked tails are horribly knotted at the tip, forming a single, dense, fleshy connection from which the entire writhing mass hangs. The abominations perform a slow, steady, hypnotic rotation on a fixed axis, performing a macabre dance within the solitary shaft of light. As the crowd turned, individual rats squirmed and struggled futilely, their microscopic movements sending constant, sickening ripples through the entire mass. The harsh light caught the damp sheen of their fur, and most terrifying of all, a thousand tiny, black, beady eyes gleamed with malevolence as they stared into the darkness, creating an atmosphere of pure, oppressive fear and visceral loathing.", + "seed": 4439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4439_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": -0.14825982379261404, + "vlm_logprob": -0.14825982379261404, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.14825982379261404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5356985330581665 + ], + "opposite_logprob_tokens": [ + -0.9106985330581665 + ], + "logprob_reward": -0.5356985330581665, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5356985330581665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1607787162065506, + 0.07166488468647003, + 0.10059788078069687, + 0.028080331161618233, + 0.08667253702878952, + 0.26759469509124756, + 0.05047803372144699, + 0.15398047864437103, + 0.028599968180060387, + 0.05155253782868385 + ], + "credit_map": [ + 0.022751962766051292, + 0.016088202595710754, + 0.022113105282187462, + 0.017920350655913353, + 0.014826448634266853, + 0.019952010363340378, + 0.018580064177513123, + 0.028546558693051338, + 0.010082975961267948, + 0.00787892285734415, + 0.01039887871593237, + 0.011146835051476955, + 0.0061439634300768375, + 0.009375481866300106, + 0.006532164756208658, + 0.010105649940669537, + 0.007597708608955145, + 0.01568901538848877, + 0.010377852246165276, + 0.01776740327477455, + 0.011801554821431637, + 0.012870049104094505, + 0.013783258385956287, + 0.010711031965911388, + 0.0024546191561967134, + 0.0030898116528987885, + 0.003270236076787114, + 0.004124889150261879, + 0.0039047575555741787, + 0.003972366917878389, + 0.003500135615468025, + 0.0037635131739079952, + 0.008404538035392761, + 0.008017639629542828, + 0.010361646302044392, + 0.017822805792093277, + 0.007077690679579973, + 0.012382502667605877, + 0.012652781791985035, + 0.009952920489013195, + 0.038833051919937134, + 0.03535676747560501, + 0.033940911293029785, + 0.03948093205690384, + 0.02554895170032978, + 0.03307776525616646, + 0.026187483221292496, + 0.03516882658004761, + 0.00650112796574831, + 0.005187428556382656, + 0.006293256767094135, + 0.008402492851018906, + 0.00486711086705327, + 0.008066078647971153, + 0.006047616712749004, + 0.005112915765494108, + 0.019447598606348038, + 0.01849861443042755, + 0.016292620450258255, + 0.019114848226308823, + 0.018977852538228035, + 0.02237035147845745, + 0.016708293929696083, + 0.022570285946130753, + 0.0036642467603087425, + 0.0022485738154500723, + 0.003965319134294987, + 0.004458094947040081, + 0.0021787129808217287, + 0.0038790337275713682, + 0.0028153855819255114, + 0.005390597973018885, + 0.004948536865413189, + 0.005650402046740055, + 0.006533598992973566, + 0.00678496528416872, + 0.008264677599072456, + 0.006479139439761639, + 0.005297307390719652, + 0.007593902759253979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.5356985330581665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033727262169122696 + ], + "opposite_logprob_tokens": [ + -3.4087271690368652 + ], + "logprob_reward": -0.033727262169122696, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033727262169122696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2079516351222992, + 0.17271558940410614, + 0.06301356106996536, + 0.05036168545484543, + 0.19018720090389252, + 0.04594786837697029, + 0.06356825679540634, + 0.13063457608222961, + 0.02470773085951805, + 0.05091197043657303 + ], + "credit_map": [ + 0.02978651225566864, + 0.025511475279927254, + 0.029856236651539803, + 0.021254533901810646, + 0.016361385583877563, + 0.03014545515179634, + 0.02152731642127037, + 0.03350870683789253, + 0.016156546771526337, + 0.01552650984376669, + 0.02677435427904129, + 0.03159961476922035, + 0.01215769536793232, + 0.028272219002246857, + 0.01890924572944641, + 0.02331940084695816, + 0.0073253256268799305, + 0.005721278488636017, + 0.008844385854899883, + 0.011471692472696304, + 0.005776883568614721, + 0.008990298956632614, + 0.00674690306186676, + 0.008136793971061707, + 0.0077146077528595924, + 0.004307513125240803, + 0.005202935077250004, + 0.007484378293156624, + 0.005553922150284052, + 0.007696985267102718, + 0.005642748437821865, + 0.006758596282452345, + 0.025908734649419785, + 0.013793149963021278, + 0.030173620209097862, + 0.02905300073325634, + 0.013821639120578766, + 0.02764957770705223, + 0.02157747559249401, + 0.028210004791617393, + 0.00555848516523838, + 0.0056393020786345005, + 0.004784449469298124, + 0.006506090518087149, + 0.005862468853592873, + 0.006491108797490597, + 0.005241249222308397, + 0.005864711944013834, + 0.006940378341823816, + 0.005016759969294071, + 0.005891468375921249, + 0.01307184062898159, + 0.005991099867969751, + 0.011821849271655083, + 0.005886610131710768, + 0.008948246017098427, + 0.018373660743236542, + 0.011897312477231026, + 0.011684093624353409, + 0.02236747369170189, + 0.013890598900616169, + 0.018851473927497864, + 0.016140712425112724, + 0.01742924377322197, + 0.0034101761411875486, + 0.0023977947421371937, + 0.002381313359364867, + 0.003804535837844014, + 0.0027875995729118586, + 0.003321403404697776, + 0.0030407514423131943, + 0.003564156126230955, + 0.005397872067987919, + 0.0046223001554608345, + 0.006146990694105625, + 0.00841765757650137, + 0.005425089504569769, + 0.007421655580401421, + 0.006781111471354961, + 0.006699294317513704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.033727262169122696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603657752275467 + ], + "opposite_logprob_tokens": [ + -1.9103658199310303 + ], + "logprob_reward": -0.1603657752275467, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603657752275467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15193824470043182, + 0.07123568654060364, + 0.11849314719438553, + 0.025616833940148354, + 0.09574959427118301, + 0.3532985746860504, + 0.029878344386816025, + 0.07664557546377182, + 0.055485907942056656, + 0.021658090874552727 + ], + "credit_map": [ + 0.01898491568863392, + 0.015729736536741257, + 0.021557701751589775, + 0.023120980709791183, + 0.011711596511304379, + 0.023525184020400047, + 0.016772301867604256, + 0.020535815507173538, + 0.009317081421613693, + 0.005073029547929764, + 0.012686440721154213, + 0.009890520945191383, + 0.005247659981250763, + 0.011328528635203838, + 0.006564248353242874, + 0.011128183454275131, + 0.016892604529857635, + 0.010144188068807125, + 0.02142932452261448, + 0.013640602119266987, + 0.008829876780509949, + 0.01817554235458374, + 0.01058406662195921, + 0.018796948716044426, + 0.0029877645429223776, + 0.0026403956580907106, + 0.0032633310183882713, + 0.003518750425428152, + 0.004238350782543421, + 0.0030542449094355106, + 0.003073523985221982, + 0.002840473083779216, + 0.00952815730124712, + 0.008461400866508484, + 0.012209027074277401, + 0.019339874386787415, + 0.007109071593731642, + 0.014366639778017998, + 0.012141603045165539, + 0.012593823485076427, + 0.04715004935860634, + 0.046509016305208206, + 0.04472251236438751, + 0.057147491723299026, + 0.03069971315562725, + 0.04451947659254074, + 0.03696276992559433, + 0.045587532222270966, + 0.0034137889742851257, + 0.003793352749198675, + 0.0023050555028021336, + 0.0060537224635481834, + 0.003957192413508892, + 0.003978550899773836, + 0.003725634189322591, + 0.002651048358529806, + 0.010409757494926453, + 0.00517121609300375, + 0.010603610426187515, + 0.009294197894632816, + 0.0061636692844331264, + 0.01184083055704832, + 0.009444118477404118, + 0.01371818408370018, + 0.0072892941534519196, + 0.0051411851309239864, + 0.007544088177382946, + 0.009965435601770878, + 0.004019328858703375, + 0.007757708430290222, + 0.004287941846996546, + 0.009480925276875496, + 0.0035599295515567064, + 0.0016436754958704114, + 0.002281867666170001, + 0.003111886791884899, + 0.0019396803108975291, + 0.0031389908399432898, + 0.002147917402908206, + 0.0038341425824910402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.1603657752275467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017176411347463727 + ], + "opposite_logprob_tokens": [ + -6.376717567443848 + ], + "logprob_reward": -0.0017176411347463727, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017176411347463727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2022588700056076, + 0.11998085677623749, + 0.06069459766149521, + 0.04387614503502846, + 0.24575799703598022, + 0.09727034717798233, + 0.0686347633600235, + 0.09442432224750519, + 0.024133339524269104, + 0.0429687462747097 + ], + "credit_map": [ + 0.02808418683707714, + 0.014872212894260883, + 0.03769746050238609, + 0.019193261861801147, + 0.022924281656742096, + 0.022629575803875923, + 0.022749895229935646, + 0.0341080017387867, + 0.013486657291650772, + 0.016675304621458054, + 0.012580584734678268, + 0.021863942965865135, + 0.012730324640870094, + 0.014733430929481983, + 0.01511934120208025, + 0.012791264802217484, + 0.007198191713541746, + 0.005700277630239725, + 0.009212390519678593, + 0.007766245864331722, + 0.00646332697942853, + 0.008465359918773174, + 0.007145906798541546, + 0.008742894977331161, + 0.00630959402769804, + 0.003956876695156097, + 0.004026835784316063, + 0.005910953972488642, + 0.008345019072294235, + 0.00517606595531106, + 0.0041811116971075535, + 0.005969690158963203, + 0.027430102229118347, + 0.020939383655786514, + 0.030763763934373856, + 0.046098243445158005, + 0.014602032490074635, + 0.0385550893843174, + 0.026799116283655167, + 0.040570277720689774, + 0.014035820960998535, + 0.013223029673099518, + 0.012229040265083313, + 0.01506438571959734, + 0.008836459368467331, + 0.01142722275108099, + 0.010221299715340137, + 0.01223309338092804, + 0.0074260421097278595, + 0.007658077869564295, + 0.004708879627287388, + 0.013915369287133217, + 0.007536354009062052, + 0.011015500873327255, + 0.007155396044254303, + 0.00921914353966713, + 0.011509958654642105, + 0.008181460201740265, + 0.011723683215677738, + 0.017003223299980164, + 0.007319842930883169, + 0.012979387305676937, + 0.012783114798367023, + 0.012923652306199074, + 0.0025052796117961407, + 0.0025849987287074327, + 0.0027686855755746365, + 0.004254033323377371, + 0.0029576821252703667, + 0.0031096553429961205, + 0.002506717573851347, + 0.003446287475526333, + 0.004621492233127356, + 0.005287632811814547, + 0.005076826550066471, + 0.005451449193060398, + 0.005685308016836643, + 0.0058983261696994305, + 0.004479100462049246, + 0.006468609906733036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0017176411347463727 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978990737348795 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.00978990737348795, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978990737348795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16343531012535095, + 0.05560657009482384, + 0.07826749980449677, + 0.03943838179111481, + 0.05237651243805885, + 0.4595009386539459, + 0.027019476518034935, + 0.04547557979822159, + 0.03269540145993233, + 0.04618431627750397 + ], + "credit_map": [ + 0.0194384902715683, + 0.014530918560922146, + 0.024194419384002686, + 0.02054634317755699, + 0.01626470498740673, + 0.01962902583181858, + 0.019704384729266167, + 0.02912701480090618, + 0.005946137011051178, + 0.006773922126740217, + 0.008402959443628788, + 0.009352593682706356, + 0.005736613180488348, + 0.007645651698112488, + 0.00493431044742465, + 0.006814376916736364, + 0.007126655895262957, + 0.009076596237719059, + 0.011769413016736507, + 0.012698542326688766, + 0.0072296313010156155, + 0.011358564719557762, + 0.00815622042864561, + 0.010851873084902763, + 0.003798969089984894, + 0.004316528327763081, + 0.004941444844007492, + 0.007921488024294376, + 0.004712417256087065, + 0.005417823791503906, + 0.004502591677010059, + 0.003827115288004279, + 0.006884701084345579, + 0.005965192802250385, + 0.006524222437292337, + 0.008341595530509949, + 0.0038540586829185486, + 0.0075955879874527454, + 0.006769717670977116, + 0.006441432982683182, + 0.06468639522790909, + 0.060817651450634, + 0.05703252553939819, + 0.07426208257675171, + 0.042266279458999634, + 0.056461088359355927, + 0.046929262578487396, + 0.05704561993479729, + 0.003970487974584103, + 0.002656316151842475, + 0.002214327920228243, + 0.005839101038873196, + 0.002315910765901208, + 0.004454048816114664, + 0.003188073867931962, + 0.002381208585575223, + 0.004820679780095816, + 0.00394309451803565, + 0.005159976426512003, + 0.007407014258205891, + 0.0048894742503762245, + 0.006847531069070101, + 0.006234946195036173, + 0.006172857712954283, + 0.004812133498489857, + 0.0025430251844227314, + 0.0038372885901480913, + 0.006447509862482548, + 0.002274701837450266, + 0.004487061873078346, + 0.0030110166408121586, + 0.0052826618775725365, + 0.004840666428208351, + 0.006737986579537392, + 0.0051124440506100655, + 0.00582589628174901, + 0.007479730527848005, + 0.005701076705008745, + 0.004469219129532576, + 0.006017289124429226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.00978990737348795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "terrifying first-person perspective shot, shot from the damp, filthy floor of a forgotten sewer, rises from an extreme low-angle perspective in an utterly terrifying manner. The camera remains still, emphasizing a sense of entrapment and helplessness. The surroundings are suffocatingly dark, their curved, leaking concrete walls coated in algae and grime in a chilling color temperature. A harsh, dusty light, streaming down from a high grate, pierces the darkness, illuminating the scene's horrific centerpiece: a grotesque, living chandelier composed of countless sewer rats. Their damp, shaggy fur—a sickening gray and brown—clings to their writhing bodies. Their long, naked tails are horribly knotted at the tip, forming a single, dense, fleshy connection from which the entire writhing mass hangs. The abominations perform a slow, steady, hypnotic rotation on a fixed axis, performing a macabre dance within the solitary shaft of light. As the crowd turned, individual rats squirmed and struggled futilely, their microscopic movements sending constant, sickening ripples through the entire mass. The harsh light caught the damp sheen of their fur, and most terrifying of all, a thousand tiny, black, beady eyes gleamed with malevolence as they stared into the darkness, creating an atmosphere of pure, oppressive fear and visceral loathing.", + "seed": 4440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_7/index_6200_seed_4440_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": -0.05859784496715292, + "vlm_logprob": -0.05859784496715292, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.05859784496715292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22356688976287842 + ], + "opposite_logprob_tokens": [ + -1.7235668897628784 + ], + "logprob_reward": -0.22356688976287842, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22356688976287842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13361060619354248, + 0.03196006268262863, + 0.03830072656273842, + 0.03239733353257179, + 0.02615368738770485, + 0.6287639141082764, + 0.06888392567634583, + 0.01129134465008974, + 0.01590478979051113, + 0.012733719311654568 + ], + "credit_map": [ + 0.009778041392564774, + 0.013790999539196491, + 0.03773002326488495, + 0.013231270015239716, + 0.007636683527380228, + 0.00777135556563735, + 0.015767350792884827, + 0.027904853224754333, + 0.0030020407866686583, + 0.003353608539327979, + 0.00535242073237896, + 0.0031105808448046446, + 0.003759878920391202, + 0.0036089776549488306, + 0.005133916158229113, + 0.0046386378817260265, + 0.0021537248976528645, + 0.0037105558440089226, + 0.00607768539339304, + 0.004604093264788389, + 0.005046616308391094, + 0.0017492322949692607, + 0.007800048217177391, + 0.007158768363296986, + 0.0018046869663521647, + 0.0012859039707109332, + 0.010336828418076038, + 0.0027652482967823744, + 0.0014433583710342646, + 0.0016394673148170114, + 0.001815883326344192, + 0.01130595337599516, + 0.0025432095862925053, + 0.002999684540554881, + 0.0037393749225884676, + 0.0028606276027858257, + 0.003906802739948034, + 0.0031461953185498714, + 0.003843477461487055, + 0.0031143114902079105, + 0.03204520419239998, + 0.0161757655441761, + 0.19615262746810913, + 0.04788118600845337, + 0.012438923120498657, + 0.034967150539159775, + 0.026454556733369827, + 0.2626483738422394, + 0.00344849843531847, + 0.003655829932540655, + 0.02296564169228077, + 0.004595979116857052, + 0.0025326977483928204, + 0.003945196978747845, + 0.00268945493735373, + 0.02505061961710453, + 0.0012418021215125918, + 0.000677623727824539, + 0.003161767730489373, + 0.0012234710156917572, + 0.0008514685905538499, + 0.0009581486228853464, + 0.0008623323519714177, + 0.002314729616045952, + 0.0013718107948079705, + 0.0017314961878582835, + 0.0032489656005054712, + 0.0015541821485385299, + 0.001669405261054635, + 0.0012465885374695063, + 0.001312599633820355, + 0.0037697413936257362, + 0.0013896076707169414, + 0.0016737987753003836, + 0.0013308316702023149, + 0.0015311797615140676, + 0.0024553232360631227, + 0.0010228513274341822, + 0.0019019918981939554, + 0.0014281336916610599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.22356688976287842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561840936541557 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.012561840936541557, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561840936541557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2926527261734009, + 0.03181059658527374, + 0.10027901828289032, + 0.05251023545861244, + 0.035816825926303864, + 0.36036229133605957, + 0.07072727382183075, + 0.015630796551704407, + 0.025302913039922714, + 0.014907343313097954 + ], + "credit_map": [ + 0.025181844830513, + 0.024145694449543953, + 0.08382532000541687, + 0.026458948850631714, + 0.024037640541791916, + 0.02196350134909153, + 0.038310904055833817, + 0.04872887581586838, + 0.002658585086464882, + 0.0037978356704115868, + 0.005715313367545605, + 0.00347975455224514, + 0.003580767661333084, + 0.0030439086258411407, + 0.005558535922318697, + 0.003975897561758757, + 0.004556040745228529, + 0.004924378357827663, + 0.031982142478227615, + 0.00859626941382885, + 0.004742407705634832, + 0.004838285967707634, + 0.006070200353860855, + 0.03456929326057434, + 0.0026148795150220394, + 0.002046142937615514, + 0.016847487539052963, + 0.00458033662289381, + 0.0017900558887049556, + 0.00274652149528265, + 0.002771334955468774, + 0.01911347731947899, + 0.003293748013675213, + 0.004065595101565123, + 0.00551280751824379, + 0.0037879026494920254, + 0.004041619133204222, + 0.005733455065637827, + 0.0032863591331988573, + 0.006095339544117451, + 0.01848343387246132, + 0.010422504507005215, + 0.11071421951055527, + 0.027742378413677216, + 0.008306563831865788, + 0.02093515917658806, + 0.01583271473646164, + 0.14792531728744507, + 0.0034008377697318792, + 0.0029682815074920654, + 0.024115925654768944, + 0.005035267677158117, + 0.0022678039968013763, + 0.003184093162417412, + 0.0028849835507571697, + 0.026870083063840866, + 0.0011038538068532944, + 0.0015781973488628864, + 0.0038029931019991636, + 0.00223529408685863, + 0.0014043059200048447, + 0.0014138690894469619, + 0.0015136536676436663, + 0.0025786294136196375, + 0.0025847717188298702, + 0.00296025606803596, + 0.004186340607702732, + 0.0030470627825707197, + 0.0028010928072035313, + 0.002280962420627475, + 0.002363575389608741, + 0.0050788503140211105, + 0.002658458426594734, + 0.0016725801397114992, + 0.0018937219865620136, + 0.0011818225029855967, + 0.0017677831929177046, + 0.0015568106900900602, + 0.002349609974771738, + 0.00182655593380332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.012561840936541557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048664771020412445 + ], + "opposite_logprob_tokens": [ + -3.0486648082733154 + ], + "logprob_reward": -0.048664771020412445, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048664771020412445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1193869337439537, + 0.10298386961221695, + 0.047436442226171494, + 0.03318459540605545, + 0.047393910586833954, + 0.4757395386695862, + 0.0749567523598671, + 0.03079136088490486, + 0.029338238760828972, + 0.03878837823867798 + ], + "credit_map": [ + 0.01103270798921585, + 0.016389021649956703, + 0.015000865794718266, + 0.018944932147860527, + 0.016805870458483696, + 0.012015956453979015, + 0.016038086265325546, + 0.013159498572349548, + 0.006960551254451275, + 0.00805321242660284, + 0.02860163152217865, + 0.008879867382347584, + 0.006654007360339165, + 0.006220562849193811, + 0.011060530319809914, + 0.026553506031632423, + 0.002831187564879656, + 0.005752380471676588, + 0.010925578884780407, + 0.005702222231775522, + 0.004878438543528318, + 0.0019975134637206793, + 0.004755882080644369, + 0.010593244805932045, + 0.0013737431727349758, + 0.0009277662029489875, + 0.011162001639604568, + 0.002773887477815151, + 0.001189139555208385, + 0.0015122033655643463, + 0.001651143655180931, + 0.012594714760780334, + 0.003823406295850873, + 0.005363474600017071, + 0.006767339538782835, + 0.004898139275610447, + 0.004966414999216795, + 0.007948612794280052, + 0.00439357990399003, + 0.009232944808900356, + 0.024554017931222916, + 0.01238183956593275, + 0.1482160985469818, + 0.03622758015990257, + 0.01006216648966074, + 0.025394661352038383, + 0.020369570702314377, + 0.19853363931179047, + 0.003932562656700611, + 0.0029047965072095394, + 0.02461055852472782, + 0.0055902451276779175, + 0.002089511835947633, + 0.0049085491336882114, + 0.0034186074044555426, + 0.027501923963427544, + 0.006166211329400539, + 0.0026905813720077276, + 0.004939838778227568, + 0.003816081676632166, + 0.0025496776215732098, + 0.0040749977342784405, + 0.002792560728266835, + 0.0037614149041473866, + 0.0020953351631760597, + 0.0031038729939609766, + 0.007532239425927401, + 0.0021663529332727194, + 0.003120425855740905, + 0.002153873210772872, + 0.002506835153326392, + 0.006659305654466152, + 0.008534557186067104, + 0.004826693329960108, + 0.005100075621157885, + 0.003356419038027525, + 0.0036268120165914297, + 0.005108455661684275, + 0.00409528985619545, + 0.00414007855579257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.048664771020412445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000563224486541003 + ], + "opposite_logprob_tokens": [ + -7.500563144683838 + ], + "logprob_reward": -0.000563224486541003, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.000563224486541003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1556597203016281, + 0.04695773869752884, + 0.0605732686817646, + 0.052630435675382614, + 0.03514809161424637, + 0.2509300708770752, + 0.31334665417671204, + 0.025424176827073097, + 0.04331836476922035, + 0.01601148396730423 + ], + "credit_map": [ + 0.009242645464837551, + 0.009748904965817928, + 0.054644759744405746, + 0.013931265100836754, + 0.008460775017738342, + 0.008760377764701843, + 0.016796037554740906, + 0.03407495096325874, + 0.0030259513296186924, + 0.004975600633770227, + 0.009807475842535496, + 0.005051495973020792, + 0.0054795765317976475, + 0.003864689962938428, + 0.007335785776376724, + 0.007417156361043453, + 0.002704260405153036, + 0.0071568419225513935, + 0.015377993695437908, + 0.005522952880710363, + 0.005704016424715519, + 0.002462832722812891, + 0.006520249880850315, + 0.015124116092920303, + 0.0020283872727304697, + 0.0019791109953075647, + 0.017393343150615692, + 0.004227557685226202, + 0.0019345879554748535, + 0.0021520722657442093, + 0.0031804703176021576, + 0.019734902307391167, + 0.0029408768750727177, + 0.004024805035442114, + 0.0068207718431949615, + 0.003355433465912938, + 0.004122822545468807, + 0.0031422663014382124, + 0.004606316797435284, + 0.006134799215942621, + 0.012827342376112938, + 0.007061381358653307, + 0.07715331017971039, + 0.01952552981674671, + 0.006181567907333374, + 0.014188529923558235, + 0.011275826022028923, + 0.10271658003330231, + 0.015496504493057728, + 0.012484746985137463, + 0.11068125814199448, + 0.020303085446357727, + 0.008933653123676777, + 0.013745960779488087, + 0.01088578812777996, + 0.12081562727689743, + 0.002012047450989485, + 0.0020986294839531183, + 0.005848333705216646, + 0.0029217542614787817, + 0.0020435985643416643, + 0.002321086125448346, + 0.003067134413868189, + 0.005111591424793005, + 0.0030356026254594326, + 0.0030337197240442038, + 0.012240803800523281, + 0.003291602013632655, + 0.0029659741558134556, + 0.003533373586833477, + 0.0026820541825145483, + 0.01253522839397192, + 0.0022850302048027515, + 0.0027463939040899277, + 0.0017385114915668964, + 0.0016478047473356128, + 0.0022487782407552004, + 0.0013853255659341812, + 0.0023685593623667955, + 0.0015910790534690022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.000563224486541003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extreme low-angle perspective, rising from the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a dark sewer with curved, leaking concrete walls coated in algae and grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a harsh, dusty light streaming down from a high grate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the centerpiece a grotesque, living chandelier composed of countless sewer rats with damp, shaggy gray and brown fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the rat mass performing a slow, steady rotation on a fixed axis while individual rats squirm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632498629391193 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632498629391193, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632498629391193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14690731465816498, + 0.038165133446455, + 0.028256572782993317, + 0.08388129621744156, + 0.022809665650129318, + 0.13983315229415894, + 0.4800325334072113, + 0.02959086373448372, + 0.012004704214632511, + 0.018518712371587753 + ], + "credit_map": [ + 0.009243853390216827, + 0.009257308207452297, + 0.052105460315942764, + 0.013890859670937061, + 0.006725828628987074, + 0.00738351559266448, + 0.015734465792775154, + 0.03256601840257645, + 0.0033948535565286875, + 0.004156356677412987, + 0.007451797369867563, + 0.003716314909979701, + 0.004012278281152248, + 0.003922289237380028, + 0.005781861022114754, + 0.00572937959805131, + 0.0017585721798241138, + 0.0022001173347234726, + 0.006366392131894827, + 0.0025897477753460407, + 0.0028694206848740578, + 0.001889824285171926, + 0.0034297893289476633, + 0.007152709178626537, + 0.0036598765291273594, + 0.0020466370042413473, + 0.029687801375985146, + 0.006732698064297438, + 0.0017877219943329692, + 0.0031849355436861515, + 0.00348000880330801, + 0.033301617950201035, + 0.0017556084785610437, + 0.0027936208061873913, + 0.004405403975397348, + 0.002207305748015642, + 0.0030979050789028406, + 0.0015061142621561885, + 0.002896585501730442, + 0.004147120751440525, + 0.008071014657616615, + 0.005866939667612314, + 0.04096926003694534, + 0.01070284191519022, + 0.004746108781546354, + 0.008485086262226105, + 0.007091842591762543, + 0.05390005186200142, + 0.02386513166129589, + 0.01882879249751568, + 0.1696571260690689, + 0.03144221007823944, + 0.013754887506365776, + 0.020450610667467117, + 0.016578566282987595, + 0.18545518815517426, + 0.003424451919272542, + 0.0030983220785856247, + 0.004343654960393906, + 0.004652517382055521, + 0.0036187253426760435, + 0.002828434109687805, + 0.0035453373566269875, + 0.004079420585185289, + 0.0008361217333003879, + 0.0009873651433736086, + 0.0027574701234698296, + 0.0011851636227220297, + 0.0012128603411838412, + 0.0009664573008194566, + 0.0009129989193752408, + 0.0031462665647268295, + 0.002327805617824197, + 0.002658698009327054, + 0.0021131967660039663, + 0.0019768276251852512, + 0.002919215941801667, + 0.00159063923638314, + 0.0029457551427185535, + 0.001986574148759246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.007632498629391193 + } + ], + "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_8/index_tensor([551, 551])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_8/index_tensor([551, 551])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b1aaea7134e5f564162b1e6223b1e822cbe8208c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_8/index_tensor([551, 551])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focused on a microphone in the center of a vast, empty stage. A single, cold-white spotlight shone vertically from above, creating a bright circle of light on the dark, polished wooden floor, while the rest of the stage was shrouded in the deep shadows of a thick velvet curtain. The central element was a classic dynamic microphone, its spherical grille and cylindrical body crafted from highly polished chrome metal, reflecting the cool light. It was barely secured by a plastic clamp atop a slender black stand. As the clamp loosened, the microphone began to slowly and uncontrollably tilt downward, its smooth metal surface rubbing against the plastic clamp. Finally, it broke free from its restraint and plummeted, driven by gravity. It struck the polished wooden floor with a resounding, resounding thud, amplified by the silence. It then bounced, rolled several times, and finally came to rest at the edge of the circle of light. The entire scene was filled with intense contrasts of light and dark, creating a dramatic atmosphere of utter tranquility and solitude, momentarily shattered.", + "seed": 16299, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_8/index_551_seed_16299_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": -0.023291715327650307, + "vlm_logprob": -0.023291715327650307, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.023291715327650307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056844864040613174 + ], + "opposite_logprob_tokens": [ + -3.43184494972229 + ], + "logprob_reward": -0.056844864040613174, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.056844864040613174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5476787090301514, + 0.10317136347293854, + 0.028225164860486984, + 0.015383766032755375, + 0.06177926063537598, + 0.029976099729537964, + 0.13077080249786377, + 0.042761411517858505, + 0.021639613434672356, + 0.01861375756561756 + ], + "credit_map": [ + 0.042927321046590805, + 0.029396414756774902, + 0.186315655708313, + 0.01334538497030735, + 0.03509689122438431, + 0.1954808235168457, + 0.014218542724847794, + 0.03089768812060356, + 0.0074652861803770065, + 0.007658890914171934, + 0.0329335518181324, + 0.0036111839581280947, + 0.006499833427369595, + 0.035413920879364014, + 0.0026695046108216047, + 0.006919194478541613, + 0.0031593297608196735, + 0.003264309372752905, + 0.006018615327775478, + 0.0015759411035105586, + 0.002697584219276905, + 0.006682618986815214, + 0.001750598312355578, + 0.003076170338317752, + 0.0015310023445636034, + 0.0018636381719261408, + 0.0027918051928281784, + 0.0013740952126681805, + 0.0015155925648286939, + 0.0029744780622422695, + 0.0012657282641157508, + 0.002067427383735776, + 0.005657898727804422, + 0.004181364551186562, + 0.019046206027269363, + 0.002697187475860119, + 0.0036462123971432447, + 0.016777830198407173, + 0.0050044902600348, + 0.004768072627484798, + 0.002095026196911931, + 0.0019384659826755524, + 0.008994814939796925, + 0.002139914548024535, + 0.0014959669206291437, + 0.00841355137526989, + 0.0024514354299753904, + 0.0024469266645610332, + 0.014029266312718391, + 0.013663629069924355, + 0.0234273299574852, + 0.01256526168435812, + 0.010368485003709793, + 0.022906724363565445, + 0.008080986328423023, + 0.025729132816195488, + 0.002303442219272256, + 0.004674146883189678, + 0.010995936580002308, + 0.002051921095699072, + 0.004898573737591505, + 0.012180330231785774, + 0.0022616868373006582, + 0.003395373933017254, + 0.0026191885117441416, + 0.0031754865776747465, + 0.0034336273092776537, + 0.0017937477678060532, + 0.002720699179917574, + 0.003266559913754463, + 0.0015001812716946006, + 0.003130124183371663, + 0.0013844359200447798, + 0.0016806625062599778, + 0.005651638377457857, + 0.0010592695325613022, + 0.0014663420151919127, + 0.005342870019376278, + 0.0006256665801629424, + 0.0014028748264536262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.056844864040613174 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609358936548233 + ], + "opposite_logprob_tokens": [ + -4.141093730926514 + ], + "logprob_reward": -0.01609358936548233, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609358936548233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7091826796531677, + 0.05448522791266441, + 0.00915438961237669, + 0.024581991136074066, + 0.011453433893620968, + 0.020580200478434563, + 0.023864684626460075, + 0.010631128214299679, + 0.030661756172776222, + 0.10540454089641571 + ], + "credit_map": [ + 0.05621090903878212, + 0.042882394045591354, + 0.2332252562046051, + 0.01934712566435337, + 0.04919968545436859, + 0.2471611201763153, + 0.019800780341029167, + 0.041355352848768234, + 0.007213323377072811, + 0.008038376457989216, + 0.009952428750693798, + 0.003560985205695033, + 0.006526790093630552, + 0.007958201691508293, + 0.0033132911194115877, + 0.007921826094388962, + 0.0009070236701518297, + 0.0012658603955060244, + 0.001454095938242972, + 0.000627768284175545, + 0.0012316599022597075, + 0.001654219115152955, + 0.0006028746720403433, + 0.001410886412486434, + 0.0023872058372944593, + 0.001576834125444293, + 0.006977217271924019, + 0.0014994916273280978, + 0.0018786549335345626, + 0.006689865607768297, + 0.0016086832620203495, + 0.001964037772268057, + 0.0009369649342261255, + 0.0007954476168379188, + 0.002959713339805603, + 0.0008529149345122278, + 0.0006906527560204268, + 0.0024651384446769953, + 0.001132691977545619, + 0.0016199098899960518, + 0.001313305925577879, + 0.0009144596988335252, + 0.006895415019243956, + 0.001292821834795177, + 0.0007129706209525466, + 0.006595974788069725, + 0.0014644595794379711, + 0.0013907916145399213, + 0.0018656793981790543, + 0.0029332260601222515, + 0.005067132413387299, + 0.0023251299280673265, + 0.00279301218688488, + 0.0048442319966852665, + 0.0011424662079662085, + 0.0028938050381839275, + 0.0009350691107101738, + 0.0015670909779146314, + 0.0020560959819704294, + 0.0011891921749338508, + 0.0013003696221858263, + 0.0017155199311673641, + 0.0006556831067427993, + 0.0012121066683903337, + 0.0033874905202537775, + 0.004765761084854603, + 0.004635835997760296, + 0.0024637053720653057, + 0.004106074571609497, + 0.005441263318061829, + 0.0014914259081706405, + 0.004370196722447872, + 0.005284544546157122, + 0.003740299493074417, + 0.03871658816933632, + 0.004329012241214514, + 0.0026131453923881054, + 0.04128057137131691, + 0.0027118860743939877, + 0.0067284852266311646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01609358936548233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347937062382698 + ], + "opposite_logprob_tokens": [ + -3.6513478755950928 + ], + "logprob_reward": -0.026347937062382698, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347937062382698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5465494990348816, + 0.02606188878417015, + 0.022786155343055725, + 0.0326235368847847, + 0.019130373373627663, + 0.02871537208557129, + 0.12947432696819305, + 0.05172252655029297, + 0.02931267023086548, + 0.11362359672784805 + ], + "credit_map": [ + 0.04194530472159386, + 0.03022666461765766, + 0.18507936596870422, + 0.013781297020614147, + 0.035007376223802567, + 0.1968332678079605, + 0.014162765815854073, + 0.029513485729694366, + 0.003864738391712308, + 0.004186471924185753, + 0.0044272225350141525, + 0.001679696375504136, + 0.0030833249911665916, + 0.004281675908714533, + 0.001339184702374041, + 0.0031995761673897505, + 0.0023491811007261276, + 0.0015720967203378677, + 0.0049519105814397335, + 0.0012066119816154242, + 0.002346295164898038, + 0.006549844518303871, + 0.0015819893451407552, + 0.0022282253485172987, + 0.0038420481141656637, + 0.003139645792543888, + 0.005479131825268269, + 0.0032503034453839064, + 0.0033121570013463497, + 0.005726818926632404, + 0.0031559302005916834, + 0.0047175027430057526, + 0.0016548522980883718, + 0.001081824884749949, + 0.005431436002254486, + 0.0015245783142745495, + 0.0009232593001797795, + 0.004963321145623922, + 0.0017960603581741452, + 0.0017550402553752065, + 0.002380827907472849, + 0.003627309575676918, + 0.006086480338126421, + 0.002295112470164895, + 0.0028275775257498026, + 0.006257190834730864, + 0.0021432056091725826, + 0.00309766735881567, + 0.012689963914453983, + 0.013101793825626373, + 0.01857908070087433, + 0.01282439474016428, + 0.009241314604878426, + 0.017861690372228622, + 0.014519347809255123, + 0.030656738206744194, + 0.0027557772118598223, + 0.00246810563839972, + 0.019024407491087914, + 0.0013707975158467889, + 0.0031625365372747183, + 0.019642645493149757, + 0.0011439248919487, + 0.002154327929019928, + 0.002492927247658372, + 0.003631850704550743, + 0.005704549141228199, + 0.0018600433832034469, + 0.00508000748232007, + 0.006434181705117226, + 0.0016788097564131021, + 0.002430301159620285, + 0.005967528093606234, + 0.0049042184837162495, + 0.042908091098070145, + 0.003794031450524926, + 0.00296258507296443, + 0.045444056391716, + 0.0025757234543561935, + 0.005067361053079367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.026347937062382698 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.4204195439815521, + 0.17347931861877441, + 0.01682285964488983, + 0.04215052351355553, + 0.031437892466783524, + 0.021779974922537804, + 0.1332414597272873, + 0.032081764191389084, + 0.039996352046728134, + 0.08859030902385712 + ], + "credit_map": [ + 0.034969788044691086, + 0.02677181549370289, + 0.13538408279418945, + 0.012293645180761814, + 0.0312897153198719, + 0.14090673625469208, + 0.013167385011911392, + 0.025636373087763786, + 0.021092943847179413, + 0.024729793891310692, + 0.032191094011068344, + 0.011178422719240189, + 0.020716270431876183, + 0.03112540766596794, + 0.009201577864587307, + 0.023243818432092667, + 0.0020323072094470263, + 0.001328836428001523, + 0.0029552665073424578, + 0.0011553465155884624, + 0.0018728849245235324, + 0.0043158563785254955, + 0.00149542058352381, + 0.0016669410979375243, + 0.003903787350282073, + 0.003061087103560567, + 0.009179109707474709, + 0.0036304774694144726, + 0.0033840343821793795, + 0.009178269654512405, + 0.0030472383368760347, + 0.00676652230322361, + 0.003384483279660344, + 0.002515177708119154, + 0.00858237687498331, + 0.0017133215442299843, + 0.0018501919694244862, + 0.007292142603546381, + 0.0029952251352369785, + 0.003104971954599023, + 0.002108177402988076, + 0.002743927761912346, + 0.003984085749834776, + 0.002526010386645794, + 0.0033895072992891073, + 0.003160639898851514, + 0.0017287955852225423, + 0.00213883095420897, + 0.009937236085534096, + 0.015623359009623528, + 0.022837605327367783, + 0.014599720016121864, + 0.011336976662278175, + 0.021827828139066696, + 0.015571702271699905, + 0.02150702476501465, + 0.0022040451876819134, + 0.0034695893991738558, + 0.009207489900290966, + 0.0015767134027555585, + 0.0036472890060395002, + 0.008883967995643616, + 0.0015117122093215585, + 0.0015809567412361503, + 0.002567748073488474, + 0.003030134132131934, + 0.011471372097730637, + 0.0024732807651162148, + 0.0027221005875617266, + 0.013141831383109093, + 0.0012172461720183492, + 0.0033726366236805916, + 0.004652988165616989, + 0.002531225560232997, + 0.03223249316215515, + 0.003789848182350397, + 0.0028371827211230993, + 0.03445867821574211, + 0.0021794713102281094, + 0.005908415652811527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "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 camera shot a close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.5799245238304138, + 0.10123015940189362, + 0.040473416447639465, + 0.03889971226453781, + 0.031764570623636246, + 0.028057660907506943, + 0.054764531552791595, + 0.06453201174736023, + 0.03099861741065979, + 0.029354853555560112 + ], + "credit_map": [ + 0.045655276626348495, + 0.03139262646436691, + 0.19629020988941193, + 0.014903663657605648, + 0.036724384874105453, + 0.20816153287887573, + 0.015681039541959763, + 0.031115775927901268, + 0.012595986016094685, + 0.015539850108325481, + 0.019048379734158516, + 0.006342877168208361, + 0.01234985701739788, + 0.01619672030210495, + 0.005595996975898743, + 0.013560488820075989, + 0.004025535192340612, + 0.00675004068762064, + 0.007838323712348938, + 0.0026215887628495693, + 0.004672824405133724, + 0.0075849853456020355, + 0.0024953484535217285, + 0.004484771750867367, + 0.002977037103846669, + 0.002991338260471821, + 0.009385461919009686, + 0.003769064089283347, + 0.00297945155762136, + 0.009677490219473839, + 0.0021681732032448053, + 0.004951696377247572, + 0.002726109931245446, + 0.0035396814346313477, + 0.008084219880402088, + 0.0014786649262532592, + 0.0032748335506767035, + 0.007099981419742107, + 0.0029430824797600508, + 0.002617996884509921, + 0.002217399189248681, + 0.0020327179227024317, + 0.00742894783616066, + 0.0020884585101157427, + 0.0016692307544872165, + 0.00725927110761404, + 0.0028293044306337833, + 0.002532330807298422, + 0.0062291608192026615, + 0.008130366913974285, + 0.006510831881314516, + 0.005006223917007446, + 0.005483061075210571, + 0.007201361935585737, + 0.00346108409576118, + 0.012742441147565842, + 0.004898144863545895, + 0.009540016762912273, + 0.011237725615501404, + 0.005003289319574833, + 0.011016055941581726, + 0.0112829077988863, + 0.004706819076091051, + 0.006847054231911898, + 0.002380321267992258, + 0.003119522938504815, + 0.008009892888367176, + 0.0020073370542377234, + 0.003924523945897818, + 0.0075212945230305195, + 0.0013561301166191697, + 0.0026795919984579086, + 0.00261361850425601, + 0.0030967784114181995, + 0.005225961096584797, + 0.0031881402246654034, + 0.004241473041474819, + 0.004274354781955481, + 0.0013667469611391425, + 0.005347779486328363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012537001632153988 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focused on a microphone in the center of a vast, empty stage. A single, cold-white spotlight shone vertically from above, creating a bright circle of light on the dark, polished wooden floor, while the rest of the stage was shrouded in the deep shadows of a thick velvet curtain. The central element was a classic dynamic microphone, its spherical grille and cylindrical body crafted from highly polished chrome metal, reflecting the cool light. It was barely secured by a plastic clamp atop a slender black stand. As the clamp loosened, the microphone began to slowly and uncontrollably tilt downward, its smooth metal surface rubbing against the plastic clamp. Finally, it broke free from its restraint and plummeted, driven by gravity. It struck the polished wooden floor with a resounding, resounding thud, amplified by the silence. It then bounced, rolled several times, and finally came to rest at the edge of the circle of light. The entire scene was filled with intense contrasts of light and dark, creating a dramatic atmosphere of utter tranquility and solitude, momentarily shattered.", + "seed": 16300, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_8/index_551_seed_16300_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": -0.02345297853462398, + "vlm_logprob": -0.02345297853462398, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.02345297853462398, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06111132726073265 + ], + "opposite_logprob_tokens": [ + -3.1861112117767334 + ], + "logprob_reward": -0.06111132726073265, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06111132726073265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09774582833051682, + 0.057975754141807556, + 0.07084015011787415, + 0.11229926347732544, + 0.07476203143596649, + 0.06088379770517349, + 0.09639435261487961, + 0.12480569630861282, + 0.20880714058876038, + 0.09548591822385788 + ], + "credit_map": [ + 0.016618119552731514, + 0.0132347596809268, + 0.01505164336413145, + 0.009531626477837563, + 0.009856067597866058, + 0.018754953518509865, + 0.006926573347300291, + 0.007772102486342192, + 0.006127504166215658, + 0.005286976229399443, + 0.010624879039824009, + 0.005190841853618622, + 0.009750581346452236, + 0.008868975564837456, + 0.00551900640130043, + 0.006606994662433863, + 0.008196866139769554, + 0.007667366415262222, + 0.01035374216735363, + 0.0059347073547542095, + 0.011333076283335686, + 0.011736871674656868, + 0.007227693218737841, + 0.00838982779532671, + 0.013501442037522793, + 0.0145541001111269, + 0.01829347386956215, + 0.012476207688450813, + 0.01324206031858921, + 0.017598535865545273, + 0.00981395598500967, + 0.012819493189454079, + 0.009160314686596394, + 0.013142609968781471, + 0.00619304645806551, + 0.0059270597994327545, + 0.014025036245584488, + 0.011968933045864105, + 0.006976241711527109, + 0.007368794642388821, + 0.007593213580548763, + 0.013071618042886257, + 0.005795781966298819, + 0.005362556781619787, + 0.0048692249692976475, + 0.01007265504449606, + 0.008077281527221203, + 0.006041472777724266, + 0.0154297249391675, + 0.0130342161282897, + 0.013878877274692059, + 0.010735697112977505, + 0.006288474425673485, + 0.01336006261408329, + 0.010685039684176445, + 0.01298227347433567, + 0.016258317977190018, + 0.013396361842751503, + 0.020727498456835747, + 0.01758740469813347, + 0.008626732043921947, + 0.020366225391626358, + 0.010585296899080276, + 0.017257874831557274, + 0.020175881683826447, + 0.06414073705673218, + 0.02284293994307518, + 0.015079083852469921, + 0.02942780964076519, + 0.028400305658578873, + 0.011834454722702503, + 0.016905952244997025, + 0.009918194264173508, + 0.021275026723742485, + 0.010270937345921993, + 0.00781144667416811, + 0.007449815049767494, + 0.01471251342445612, + 0.01238769106566906, + 0.011660300195217133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.06111132726073265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008665576577186584 + ], + "opposite_logprob_tokens": [ + -4.758665561676025 + ], + "logprob_reward": -0.008665576577186584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008665576577186584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18894566595554352, + 0.20883120596408844, + 0.09361788630485535, + 0.03341994434595108, + 0.12992043793201447, + 0.07493601739406586, + 0.03419974818825722, + 0.04484698548913002, + 0.10449858009815216, + 0.08678343147039413 + ], + "credit_map": [ + 0.020710503682494164, + 0.0325847826898098, + 0.021943001076579094, + 0.013774847611784935, + 0.032294437289237976, + 0.02926216833293438, + 0.023361990228295326, + 0.015013927593827248, + 0.02506008744239807, + 0.02332204207777977, + 0.019720952957868576, + 0.013741839677095413, + 0.04849296435713768, + 0.03406788781285286, + 0.026802336797118187, + 0.01762310601770878, + 0.01036111917346716, + 0.009659544564783573, + 0.010813113301992416, + 0.006937365047633648, + 0.02156013995409012, + 0.015667708590626717, + 0.009499281644821167, + 0.009119613096117973, + 0.003899357980117202, + 0.0031490721739828587, + 0.005791312549263239, + 0.003330262377858162, + 0.005622330587357283, + 0.004300240892916918, + 0.003456337610259652, + 0.003871029708534479, + 0.014419609680771828, + 0.018083980306982994, + 0.013011543080210686, + 0.006966734305024147, + 0.031410448253154755, + 0.021372800692915916, + 0.015439635142683983, + 0.00921568926423788, + 0.007279734127223492, + 0.011039665900170803, + 0.006495242472738028, + 0.007346304599195719, + 0.006494882516562939, + 0.010776428505778313, + 0.017232241109013557, + 0.008271520957350731, + 0.004424817860126495, + 0.005015700589865446, + 0.005201234947890043, + 0.0034354578237980604, + 0.0026072945911437273, + 0.005124412477016449, + 0.00373280537314713, + 0.00465802289545536, + 0.006448790896683931, + 0.006614238489419222, + 0.0070191435515880585, + 0.005176724400371313, + 0.003199636237695813, + 0.006588146556168795, + 0.003532260423526168, + 0.006268046796321869, + 0.01057451032102108, + 0.03577369078993797, + 0.008773219771683216, + 0.006450877990573645, + 0.01049793791025877, + 0.020157478749752045, + 0.005159453488886356, + 0.007111406419426203, + 0.010619222186505795, + 0.022214213386178017, + 0.008928791619837284, + 0.00699193961918354, + 0.011802665889263153, + 0.011203999631106853, + 0.006787830498069525, + 0.008234772831201553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.008665576577186584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329319901764393 + ], + "opposite_logprob_tokens": [ + -3.7732932567596436 + ], + "logprob_reward": -0.02329319901764393, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329319901764393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09852664917707443, + 0.052404604852199554, + 0.1129414364695549, + 0.10755693912506104, + 0.30091339349746704, + 0.20775745809078217, + 0.04247017204761505, + 0.02328454703092575, + 0.038104213774204254, + 0.016040539368987083 + ], + "credit_map": [ + 0.014444862492382526, + 0.012488432228565216, + 0.011667010374367237, + 0.007145586423575878, + 0.01930936984717846, + 0.014654348604381084, + 0.009664208628237247, + 0.009152843616902828, + 0.008236199617385864, + 0.0057654301635921, + 0.005681103095412254, + 0.004074832424521446, + 0.008737819269299507, + 0.009883374907076359, + 0.005547174718230963, + 0.004478675313293934, + 0.01598750427365303, + 0.01225526724010706, + 0.01207471452653408, + 0.006931636016815901, + 0.021532386541366577, + 0.019849147647619247, + 0.014943414367735386, + 0.009367376565933228, + 0.01479804515838623, + 0.013098645024001598, + 0.010548621416091919, + 0.006878548301756382, + 0.020864732563495636, + 0.0206876490265131, + 0.011471027508378029, + 0.009209686890244484, + 0.03495050594210625, + 0.039137523621320724, + 0.030050737783312798, + 0.016534771770238876, + 0.06970609724521637, + 0.05163286253809929, + 0.03146351873874664, + 0.027437401935458183, + 0.032000184059143066, + 0.03245574235916138, + 0.020183999091386795, + 0.017484983429312706, + 0.014046825468540192, + 0.04215013235807419, + 0.028105752542614937, + 0.021329859271645546, + 0.005622470751404762, + 0.004190724343061447, + 0.003909931983798742, + 0.004565189126878977, + 0.0032648034393787384, + 0.0069478158839046955, + 0.008454186841845512, + 0.005515052936971188, + 0.0029726997017860413, + 0.004874961916357279, + 0.0034534370061010122, + 0.002018583472818136, + 0.0019508077530190349, + 0.0033999120350927114, + 0.0025154887698590755, + 0.002098659286275506, + 0.0038122977130115032, + 0.011227165348827839, + 0.0042641437612473965, + 0.0028646544087678194, + 0.0037806029431521893, + 0.00600106455385685, + 0.002968753920868039, + 0.003185535315424204, + 0.0022435837890952826, + 0.0013185949064791203, + 0.001845374470576644, + 0.001670323545113206, + 0.0016193960327655077, + 0.00242312322370708, + 0.0024613174609839916, + 0.002458827802911401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.02329319901764393 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036159863229840994 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036159863229840994, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036159863229840994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15226775407791138, + 0.06663393974304199, + 0.06967941671609879, + 0.05940522626042366, + 0.10419907420873642, + 0.20617534220218658, + 0.09177897125482559, + 0.06469818204641342, + 0.06041151285171509, + 0.12475059926509857 + ], + "credit_map": [ + 0.02586250938475132, + 0.014615735039114952, + 0.020348433405160904, + 0.012184648774564266, + 0.019251832738518715, + 0.027615925297141075, + 0.01672516204416752, + 0.015663521364331245, + 0.010253874585032463, + 0.006698880810290575, + 0.008967529982328415, + 0.006260360591113567, + 0.00815520528703928, + 0.010253001004457474, + 0.007658676244318485, + 0.00838641170412302, + 0.009409351274371147, + 0.007643613964319229, + 0.009152933955192566, + 0.005660143215209246, + 0.011500129476189613, + 0.009541438892483711, + 0.009102720767259598, + 0.007669087499380112, + 0.008042998611927032, + 0.007428682409226894, + 0.006233466323465109, + 0.004660796374082565, + 0.01042802445590496, + 0.01161924283951521, + 0.006337041966617107, + 0.004654971417039633, + 0.012963171117007732, + 0.013406272046267986, + 0.010742017067968845, + 0.008197835646569729, + 0.02087464928627014, + 0.016804596409201622, + 0.011730978265404701, + 0.00947955809533596, + 0.03158204257488251, + 0.030555235221982002, + 0.019828923046588898, + 0.01749039813876152, + 0.01755543239414692, + 0.0391155481338501, + 0.03133416920900345, + 0.018713587895035744, + 0.010951164178550243, + 0.009578519500792027, + 0.012007375247776508, + 0.009677992202341557, + 0.008095470257103443, + 0.014407404698431492, + 0.017527330666780472, + 0.009533715434372425, + 0.007881826721131802, + 0.007307065650820732, + 0.0090069230645895, + 0.006104252301156521, + 0.008030395954847336, + 0.010652710683643818, + 0.007896486669778824, + 0.007818525657057762, + 0.008730728179216385, + 0.008506480604410172, + 0.009357372298836708, + 0.006828030571341515, + 0.0052298493683338165, + 0.009257780387997627, + 0.006572723854333162, + 0.005928552243858576, + 0.013406220823526382, + 0.012496564537286758, + 0.01816437765955925, + 0.01507593598216772, + 0.008703457191586494, + 0.017764439806342125, + 0.010601122863590717, + 0.028538478538393974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0036159863229840994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057880349457264 + ], + "opposite_logprob_tokens": [ + -3.8955788612365723 + ], + "logprob_reward": -0.02057880349457264, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057880349457264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20027174055576324, + 0.06770645827054977, + 0.18599998950958252, + 0.053491294384002686, + 0.07492703199386597, + 0.05565565451979637, + 0.07764046639204025, + 0.07281938195228577, + 0.08183947950601578, + 0.12964850664138794 + ], + "credit_map": [ + 0.024827172979712486, + 0.022841677069664, + 0.0344168096780777, + 0.01980038359761238, + 0.022797562181949615, + 0.03389116749167442, + 0.01784091629087925, + 0.023856045678257942, + 0.008025114424526691, + 0.009187955409288406, + 0.010031425394117832, + 0.006560761947184801, + 0.008119096048176289, + 0.011306103318929672, + 0.0065139238722622395, + 0.007962080650031567, + 0.026869362220168114, + 0.020694300532341003, + 0.032501526176929474, + 0.02416965365409851, + 0.015484238043427467, + 0.03026973456144333, + 0.010607464239001274, + 0.025403713807463646, + 0.007995903491973877, + 0.006966637913137674, + 0.007127244025468826, + 0.003602825105190277, + 0.008365881629288197, + 0.00892373826354742, + 0.0042233639396727085, + 0.006285699084401131, + 0.009103434160351753, + 0.010735145770013332, + 0.009978678077459335, + 0.006049404386430979, + 0.011923901736736298, + 0.013462714850902557, + 0.005945221055299044, + 0.007728532422333956, + 0.00639092642813921, + 0.011148504912853241, + 0.007893014699220657, + 0.005671265535056591, + 0.004123865161091089, + 0.008737738244235516, + 0.00511170132085681, + 0.0065786391496658325, + 0.010226576589047909, + 0.018511798232793808, + 0.007728183642029762, + 0.006593963131308556, + 0.006680372171103954, + 0.012247370555996895, + 0.009684962220489979, + 0.00596723472699523, + 0.007717920001596212, + 0.020928939804434776, + 0.007291123736649752, + 0.006446478888392448, + 0.006512867286801338, + 0.013345922343432903, + 0.004626589827239513, + 0.005949542857706547, + 0.008288919925689697, + 0.027631651610136032, + 0.007045848295092583, + 0.005886860657483339, + 0.007689358666539192, + 0.015894398093223572, + 0.0033831936307251453, + 0.006019254680722952, + 0.015774110332131386, + 0.038567133247852325, + 0.0109579311683774, + 0.009523999877274036, + 0.012380140833556652, + 0.022264815866947174, + 0.005428314674645662, + 0.01475206483155489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02057880349457264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focused on a microphone in the center of a vast, empty stage. A single, cold-white spotlight shone vertically from above, creating a bright circle of light on the dark, polished wooden floor, while the rest of the stage was shrouded in the deep shadows of a thick velvet curtain. The central element was a classic dynamic microphone, its spherical grille and cylindrical body crafted from highly polished chrome metal, reflecting the cool light. It was barely secured by a plastic clamp atop a slender black stand. As the clamp loosened, the microphone began to slowly and uncontrollably tilt downward, its smooth metal surface rubbing against the plastic clamp. Finally, it broke free from its restraint and plummeted, driven by gravity. It struck the polished wooden floor with a resounding, resounding thud, amplified by the silence. It then bounced, rolled several times, and finally came to rest at the edge of the circle of light. The entire scene was filled with intense contrasts of light and dark, creating a dramatic atmosphere of utter tranquility and solitude, momentarily shattered.", + "seed": 16301, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_8/index_551_seed_16301_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": -0.02050355188548565, + "vlm_logprob": -0.02050355188548565, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02050355188548565, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05623302608728409 + ], + "opposite_logprob_tokens": [ + -3.3062329292297363 + ], + "logprob_reward": -0.05623302608728409, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05623302608728409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1383417844772339, + 0.08225070685148239, + 0.08728861063718796, + 0.06963422894477844, + 0.1078694760799408, + 0.08215082436800003, + 0.06137768179178238, + 0.25316116213798523, + 0.06353382766246796, + 0.05439168959856033 + ], + "credit_map": [ + 0.012691440060734749, + 0.011221677996218204, + 0.018735496327280998, + 0.015018930658698082, + 0.01586274243891239, + 0.0193528663367033, + 0.02645944617688656, + 0.018999187275767326, + 0.008359049446880817, + 0.01283282320946455, + 0.010757378302514553, + 0.012613780796527863, + 0.00748430797830224, + 0.011310243979096413, + 0.01020718738436699, + 0.008685936219990253, + 0.009418617933988571, + 0.010394209995865822, + 0.010881076566874981, + 0.014137594029307365, + 0.010184330865740776, + 0.01404652651399374, + 0.00930896308273077, + 0.008917293511331081, + 0.0074650878086686134, + 0.005922677926719189, + 0.00959907379001379, + 0.0112516600638628, + 0.0070656281895935535, + 0.013225126080214977, + 0.007598054129630327, + 0.007506921887397766, + 0.014153136871755123, + 0.01002495177090168, + 0.013674536719918251, + 0.011526672169566154, + 0.01043604500591755, + 0.022596850991249084, + 0.016693413257598877, + 0.0087638720870018, + 0.008138501085340977, + 0.012747718952596188, + 0.009064500220119953, + 0.010546080768108368, + 0.010156875476241112, + 0.011804774403572083, + 0.01025764923542738, + 0.009434725157916546, + 0.004772016778588295, + 0.008061259984970093, + 0.006394603755325079, + 0.007235171739012003, + 0.006747287232428789, + 0.010168696753680706, + 0.008033241145312786, + 0.009965404868125916, + 0.02077700011432171, + 0.05994442477822304, + 0.013755626045167446, + 0.011064421385526657, + 0.061508722603321075, + 0.054446760565042496, + 0.015520756132900715, + 0.01614343747496605, + 0.004629779141396284, + 0.012024558149278164, + 0.0059049068950116634, + 0.006852542050182819, + 0.01185937412083149, + 0.010240249335765839, + 0.00612972117960453, + 0.005892699118703604, + 0.005172082222998142, + 0.0064275264739990234, + 0.0057083782739937305, + 0.004694631788879633, + 0.008653027936816216, + 0.012067115865647793, + 0.006055229343473911, + 0.005613700486719608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.05623302608728409 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219040982425213 + ], + "opposite_logprob_tokens": [ + -4.264218807220459 + ], + "logprob_reward": -0.014219040982425213, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219040982425213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1834045946598053, + 0.07916949689388275, + 0.15079820156097412, + 0.05117425322532654, + 0.02518223039805889, + 0.13912664353847504, + 0.0892677903175354, + 0.15339215099811554, + 0.07673922926187515, + 0.05174535885453224 + ], + "credit_map": [ + 0.02043735235929489, + 0.012164892628788948, + 0.019631892442703247, + 0.019912542775273323, + 0.018079202622175217, + 0.03851713612675667, + 0.03375733271241188, + 0.02090425044298172, + 0.009754584170877934, + 0.010385465808212757, + 0.007905897684395313, + 0.008892313577234745, + 0.011255831457674503, + 0.01107401866465807, + 0.011655433103442192, + 0.008245947770774364, + 0.01634278893470764, + 0.01924767531454563, + 0.018589051440358162, + 0.020255783572793007, + 0.016963094472885132, + 0.02407785877585411, + 0.020246511325240135, + 0.015075447037816048, + 0.004584421869367361, + 0.005674035754054785, + 0.0065856808796525, + 0.00963278952986002, + 0.005009740591049194, + 0.006686826702207327, + 0.00699900696054101, + 0.006001751404255629, + 0.00327891088090837, + 0.001999309752136469, + 0.0033017443493008614, + 0.003040482522919774, + 0.0024259816855192184, + 0.004767109174281359, + 0.00394394900649786, + 0.002424742793664336, + 0.010574043728411198, + 0.02285834215581417, + 0.017734892666339874, + 0.02004874125123024, + 0.0166228748857975, + 0.019795048981904984, + 0.0172946248203516, + 0.014198086224496365, + 0.007104393560439348, + 0.013028154149651527, + 0.010578387416899204, + 0.011540967971086502, + 0.010921566747128963, + 0.015013341791927814, + 0.00942271202802658, + 0.01165825966745615, + 0.011720295995473862, + 0.03809415176510811, + 0.008196325972676277, + 0.006120072677731514, + 0.03902748227119446, + 0.03297636657953262, + 0.008623627945780754, + 0.008633818477392197, + 0.006489932537078857, + 0.01490880735218525, + 0.00594266876578331, + 0.005630067549645901, + 0.01833612471818924, + 0.012858104892075062, + 0.005749419331550598, + 0.006824100855737925, + 0.00406799279153347, + 0.006143628619611263, + 0.005236230790615082, + 0.005344698671251535, + 0.010043617337942123, + 0.00820682942867279, + 0.006948880385607481, + 0.005753479897975922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.014219040982425213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192220479249954 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192220479249954, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192220479249954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16295544803142548, + 0.09918893873691559, + 0.08927791565656662, + 0.030219605192542076, + 0.030286749824881554, + 0.055086709558963776, + 0.10390397906303406, + 0.21797437965869904, + 0.15243376791477203, + 0.05867255851626396 + ], + "credit_map": [ + 0.01845794916152954, + 0.010855590924620628, + 0.015972372144460678, + 0.015494218096137047, + 0.01900153048336506, + 0.034031160175800323, + 0.03159720450639725, + 0.017545409500598907, + 0.012812307104468346, + 0.01539541408419609, + 0.0107143335044384, + 0.014136062003672123, + 0.009521681815385818, + 0.01844840869307518, + 0.00978177972137928, + 0.0083789536729455, + 0.009149591438472271, + 0.00898421835154295, + 0.013014417141675949, + 0.018555348739027977, + 0.006509233266115189, + 0.01298991683870554, + 0.01152085792273283, + 0.008554330095648766, + 0.004174747504293919, + 0.0023424362298101187, + 0.0031277858652174473, + 0.00521653238683939, + 0.002385869389399886, + 0.005781854037195444, + 0.00399977108463645, + 0.003190606599673629, + 0.003171833697706461, + 0.0022192851174622774, + 0.004975640214979649, + 0.004375608637928963, + 0.002799644600600004, + 0.0039983512833714485, + 0.004485274665057659, + 0.004261107183992863, + 0.005411871243268251, + 0.006163625046610832, + 0.006275517866015434, + 0.006354149896651506, + 0.006652283016592264, + 0.008598517626523972, + 0.008420811034739017, + 0.007209933362901211, + 0.00789482519030571, + 0.01741199754178524, + 0.012268583290278912, + 0.013654333539307117, + 0.011543277651071548, + 0.016807392239570618, + 0.01091585773974657, + 0.013407708145678043, + 0.01770641840994358, + 0.05505538359284401, + 0.012036609463393688, + 0.011072559282183647, + 0.05298542231321335, + 0.04171299934387207, + 0.016074728220701218, + 0.011330232955515385, + 0.011416267603635788, + 0.036688875406980515, + 0.009500790387392044, + 0.014232834801077843, + 0.0217591542750597, + 0.02652778849005699, + 0.018030809238553047, + 0.014277245849370956, + 0.004865926690399647, + 0.011813431046903133, + 0.004403437487781048, + 0.003809812478721142, + 0.013350757770240307, + 0.010615765117108822, + 0.005032627377659082, + 0.004780796822160482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.018192220479249954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093126393854618 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093126393854618, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093126393854618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09856773912906647, + 0.04857555776834488, + 0.1321338713169098, + 0.03765405714511871, + 0.06493666768074036, + 0.133900985121727, + 0.027223190292716026, + 0.27255669236183167, + 0.1474180817604065, + 0.03703314810991287 + ], + "credit_map": [ + 0.010751734487712383, + 0.0062922821380198, + 0.01070986408740282, + 0.00917636975646019, + 0.011312807910144329, + 0.022812191396951675, + 0.018858101218938828, + 0.008654388599097729, + 0.006501738913357258, + 0.004874035716056824, + 0.00630768109112978, + 0.005001879297196865, + 0.003564052516594529, + 0.013310214504599571, + 0.004525850061327219, + 0.004490105900913477, + 0.011698679067194462, + 0.014947818592190742, + 0.018952785059809685, + 0.029865236952900887, + 0.011402430012822151, + 0.017301302403211594, + 0.01714385487139225, + 0.010821758769452572, + 0.004821310751140118, + 0.0037158417981117964, + 0.005223976913839579, + 0.006151951849460602, + 0.002937780227512121, + 0.005869923625141382, + 0.004695035982877016, + 0.004238235764205456, + 0.008120527490973473, + 0.004637157078832388, + 0.009889674372971058, + 0.0071164146065711975, + 0.005853673443198204, + 0.01544435229152441, + 0.007649287581443787, + 0.006225581280887127, + 0.016683176159858704, + 0.01792907901108265, + 0.017886701971292496, + 0.015338442288339138, + 0.0095085883513093, + 0.025600427761673927, + 0.011507205665111542, + 0.019447367638349533, + 0.0027687735855579376, + 0.004008139483630657, + 0.0029696524143218994, + 0.0028672555927187204, + 0.0038455072790384293, + 0.003893359797075391, + 0.0033231801353394985, + 0.003547320142388344, + 0.020654313266277313, + 0.06821341812610626, + 0.016838405281305313, + 0.01339774765074253, + 0.07027483731508255, + 0.05411447212100029, + 0.015992822125554085, + 0.013070669956505299, + 0.01191813312470913, + 0.02132248878479004, + 0.016797956079244614, + 0.016798662021756172, + 0.020836874842643738, + 0.020268486812710762, + 0.018918145447969437, + 0.020557330921292305, + 0.002988701919093728, + 0.004597718361765146, + 0.004477820359170437, + 0.002787875710055232, + 0.006631923373788595, + 0.0072309887036681175, + 0.003951626364141703, + 0.0043664914555847645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.004093126393854618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16104406118392944, + 0.08168108761310577, + 0.1933184266090393, + 0.08395793288946152, + 0.05037948116660118, + 0.10002551227807999, + 0.10610029101371765, + 0.1237001046538353, + 0.0649792030453682, + 0.034813933074474335 + ], + "credit_map": [ + 0.01585313491523266, + 0.011536207981407642, + 0.019174573943018913, + 0.014504970982670784, + 0.029729429632425308, + 0.02840128168463707, + 0.02703176811337471, + 0.014812694862484932, + 0.01041306834667921, + 0.006219374947249889, + 0.014929661527276039, + 0.007849607616662979, + 0.00848101731389761, + 0.013131018728017807, + 0.008312173187732697, + 0.012345165014266968, + 0.01687978208065033, + 0.020737675949931145, + 0.026513386517763138, + 0.030248351395130157, + 0.022504353895783424, + 0.03698036074638367, + 0.018895694985985756, + 0.02055881731212139, + 0.009954954497516155, + 0.007888748310506344, + 0.010393126867711544, + 0.015672767534852028, + 0.007317321840673685, + 0.012266192585229874, + 0.012026102282106876, + 0.008438723161816597, + 0.006766323931515217, + 0.003052926156669855, + 0.01107035018503666, + 0.005223829299211502, + 0.0038865620736032724, + 0.0086803138256073, + 0.005250985734164715, + 0.006448189727962017, + 0.011323357932269573, + 0.015216229483485222, + 0.01276552677154541, + 0.007849212735891342, + 0.009222134947776794, + 0.02397667057812214, + 0.0074808718636631966, + 0.012191510759294033, + 0.010072031058371067, + 0.015017093159258366, + 0.01434759609401226, + 0.011823588982224464, + 0.010466177947819233, + 0.013856506906449795, + 0.013624985702335835, + 0.01689230278134346, + 0.009462477639317513, + 0.031342145055532455, + 0.006788966245949268, + 0.005856603384017944, + 0.030194036662578583, + 0.026921160519123077, + 0.006416906137019396, + 0.006717815063893795, + 0.005330587271600962, + 0.011178159154951572, + 0.00599104817956686, + 0.005549747031182051, + 0.010967769660055637, + 0.009573141112923622, + 0.00767275458201766, + 0.008716003969311714, + 0.0030037215910851955, + 0.005897996015846729, + 0.003004471305757761, + 0.0022328600753098726, + 0.0066579654812812805, + 0.008076483383774757, + 0.003218255005776882, + 0.002722180914133787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.009780345484614372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focused on a microphone in the center of a vast, empty stage. A single, cold-white spotlight shone vertically from above, creating a bright circle of light on the dark, polished wooden floor, while the rest of the stage was shrouded in the deep shadows of a thick velvet curtain. The central element was a classic dynamic microphone, its spherical grille and cylindrical body crafted from highly polished chrome metal, reflecting the cool light. It was barely secured by a plastic clamp atop a slender black stand. As the clamp loosened, the microphone began to slowly and uncontrollably tilt downward, its smooth metal surface rubbing against the plastic clamp. Finally, it broke free from its restraint and plummeted, driven by gravity. It struck the polished wooden floor with a resounding, resounding thud, amplified by the silence. It then bounced, rolled several times, and finally came to rest at the edge of the circle of light. The entire scene was filled with intense contrasts of light and dark, creating a dramatic atmosphere of utter tranquility and solitude, momentarily shattered.", + "seed": 16302, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_8/index_551_seed_16302_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": -0.016892333654686807, + "vlm_logprob": -0.016892333654686807, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.016892333654686807, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027358390390872955 + ], + "opposite_logprob_tokens": [ + -5.027358531951904 + ], + "logprob_reward": -0.027358390390872955, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.027358390390872955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040177084505558014, + 0.022520016878843307, + 0.4140702188014984, + 0.0294872485101223, + 0.011622435413300991, + 0.04933843016624451, + 0.12945038080215454, + 0.06621209532022476, + 0.0537850446999073, + 0.18333709239959717 + ], + "credit_map": [ + 0.004042972810566425, + 0.006303077097982168, + 0.003964122384786606, + 0.006355633959174156, + 0.0045599183067679405, + 0.005647482816129923, + 0.005636586807668209, + 0.003667288925498724, + 0.0020917048677802086, + 0.0020837250631302595, + 0.003692381316795945, + 0.002194364322349429, + 0.0017846161499619484, + 0.0035554692149162292, + 0.002803560346364975, + 0.0043141948990523815, + 0.031422968953847885, + 0.037789490073919296, + 0.06353217363357544, + 0.10521562397480011, + 0.023613611236214638, + 0.057123661041259766, + 0.04416967183351517, + 0.051203012466430664, + 0.002694410039111972, + 0.002869914984330535, + 0.003836132353171706, + 0.004980678670108318, + 0.0026226758491247892, + 0.004603685811161995, + 0.003358106128871441, + 0.004521642811596394, + 0.0009776988299563527, + 0.0012202640064060688, + 0.0015333310002461076, + 0.0015817423118278384, + 0.0015856378013268113, + 0.0018096667481586337, + 0.0010726338950917125, + 0.0018414605874568224, + 0.005560303572565317, + 0.005488825961947441, + 0.0075673311948776245, + 0.007282925304025412, + 0.004581528250128031, + 0.007190247531980276, + 0.003909504506736994, + 0.007757765240967274, + 0.007394595071673393, + 0.035722434520721436, + 0.00694746058434248, + 0.0074225738644599915, + 0.019837509840726852, + 0.03359387814998627, + 0.010518052615225315, + 0.008013890124857426, + 0.002870434895157814, + 0.021712010726332664, + 0.0031694057397544384, + 0.003260147524997592, + 0.005202523432672024, + 0.021614419296383858, + 0.005208808928728104, + 0.0031743452418595552, + 0.0032707368955016136, + 0.012551242485642433, + 0.0023228065110743046, + 0.0034030727110803127, + 0.009176516905426979, + 0.014185277745127678, + 0.00570912379771471, + 0.0031662671826779842, + 0.008420600555837154, + 0.0671677216887474, + 0.005090615712106228, + 0.005782942287623882, + 0.03189215436577797, + 0.05414976924657822, + 0.005013668444007635, + 0.005819621030241251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.027358390390872955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036353471223264933 + ], + "opposite_logprob_tokens": [ + -5.628635406494141 + ], + "logprob_reward": -0.0036353471223264933, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036353471223264933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06296553462743759, + 0.03198377788066864, + 0.09713605046272278, + 0.08114858716726303, + 0.04903094097971916, + 0.04988046735525131, + 0.1755463033914566, + 0.1222563087940216, + 0.11097125709056854, + 0.21908079087734222 + ], + "credit_map": [ + 0.006963686551898718, + 0.007990087382495403, + 0.007970461621880531, + 0.007391572464257479, + 0.0073898253031075, + 0.008377375081181526, + 0.007281978148967028, + 0.009600553661584854, + 0.0030568353831768036, + 0.004800601862370968, + 0.004517121706157923, + 0.0033117984421551228, + 0.003684791037812829, + 0.004733129870146513, + 0.0033115465193986893, + 0.004567952826619148, + 0.00843952875584364, + 0.00980774313211441, + 0.014928081072866917, + 0.021519077941775322, + 0.006980360019952059, + 0.012747281230986118, + 0.008223348297178745, + 0.014490630477666855, + 0.007881046272814274, + 0.011559839360415936, + 0.01105438731610775, + 0.008290070109069347, + 0.011458846740424633, + 0.012058094143867493, + 0.00751901650801301, + 0.011327282525599003, + 0.004099010955542326, + 0.004481135867536068, + 0.005965884309262037, + 0.008386091329157352, + 0.005182876251637936, + 0.007317529059946537, + 0.005124417133629322, + 0.00847399327903986, + 0.006144416518509388, + 0.005096351262181997, + 0.007005851250141859, + 0.007445655297487974, + 0.00455322302877903, + 0.006720442790538073, + 0.004292817320674658, + 0.00862171035259962, + 0.009811101481318474, + 0.04943803697824478, + 0.00916683953255415, + 0.011190027929842472, + 0.027063630521297455, + 0.04685373976826668, + 0.011171295307576656, + 0.010851633735001087, + 0.0055352249182760715, + 0.046063363552093506, + 0.0053438651375472546, + 0.0046749054454267025, + 0.010443005710840225, + 0.039396196603775024, + 0.004473611246794462, + 0.006326138973236084, + 0.005617342423647642, + 0.026126345619559288, + 0.005054358392953873, + 0.006314859725534916, + 0.019398406147956848, + 0.031956665217876434, + 0.008746990002691746, + 0.007756300736218691, + 0.007667422294616699, + 0.08265585452318192, + 0.006443483754992485, + 0.004457678180187941, + 0.03151123225688934, + 0.07457499951124191, + 0.004896542057394981, + 0.0068735782988369465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.0036353471223264933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542063370347023 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542063370347023, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542063370347023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0246859360486269, + 0.024768682196736336, + 0.06345567107200623, + 0.1031026691198349, + 0.02795940637588501, + 0.057227443903684616, + 0.08392354100942612, + 0.039088159799575806, + 0.43700310587882996, + 0.13878542184829712 + ], + "credit_map": [ + 0.002730682725086808, + 0.0022668326273560524, + 0.004042688757181168, + 0.0027999000158160925, + 0.0024109359364956617, + 0.003979887813329697, + 0.0028320979326963425, + 0.003622907679527998, + 0.0024125229101628065, + 0.001721134758554399, + 0.004231846891343594, + 0.0025033485144376755, + 0.002013523131608963, + 0.003981884103268385, + 0.0034812979865819216, + 0.004423120524734259, + 0.004658128134906292, + 0.006745221558958292, + 0.008502609096467495, + 0.01633027382194996, + 0.004095897544175386, + 0.008184642530977726, + 0.007619339041411877, + 0.007319555152207613, + 0.00799738522619009, + 0.013461140915751457, + 0.01901126839220524, + 0.01540924422442913, + 0.009303837083280087, + 0.015300040133297443, + 0.008257914334535599, + 0.014361826702952385, + 0.002609900664538145, + 0.0023406841792166233, + 0.0030255874153226614, + 0.004487015772610903, + 0.0031532705761492252, + 0.004403588827699423, + 0.003444008296355605, + 0.004495348781347275, + 0.006503669079393148, + 0.005466084461659193, + 0.009383894503116608, + 0.008492860943078995, + 0.004059086553752422, + 0.008594942279160023, + 0.004373154602944851, + 0.010353744961321354, + 0.007530242204666138, + 0.011135784909129143, + 0.004936561919748783, + 0.00938475877046585, + 0.017644688487052917, + 0.012115860357880592, + 0.011076565831899643, + 0.010099068284034729, + 0.003044458804652095, + 0.007463664282113314, + 0.0038950524758547544, + 0.003988702781498432, + 0.005116793792694807, + 0.006756487302482128, + 0.005192684940993786, + 0.003630310297012329, + 0.019074970856308937, + 0.12757456302642822, + 0.010837579146027565, + 0.017597513273358345, + 0.11008008569478989, + 0.10738246887922287, + 0.028376363217830658, + 0.016079513356089592, + 0.009531775489449501, + 0.047585416585206985, + 0.004562340211123228, + 0.005131927318871021, + 0.02057592198252678, + 0.04034937918186188, + 0.004313548561185598, + 0.006735100876539946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.012542063370347023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03771159052848816, + 0.009613682515919209, + 0.12886324524879456, + 0.020567843690514565, + 0.016022732481360435, + 0.022950835525989532, + 0.07305850088596344, + 0.0538996197283268, + 0.15368889272212982, + 0.48362302780151367 + ], + "credit_map": [ + 0.0035842813085764647, + 0.005589304957538843, + 0.004696988966315985, + 0.005343518685549498, + 0.0036803761031478643, + 0.005703976843506098, + 0.0039842501282691956, + 0.005128891672939062, + 0.0008696021977812052, + 0.001103996648453176, + 0.001323057571426034, + 0.0010312962112948298, + 0.0009357034577988088, + 0.0014234973350539804, + 0.0013833169359713793, + 0.0015432119835168123, + 0.009499768726527691, + 0.01396031305193901, + 0.02422911487519741, + 0.02683105133473873, + 0.009262300096452236, + 0.018488649278879166, + 0.012671117670834064, + 0.013920934870839119, + 0.0018391936318948865, + 0.0034178276546299458, + 0.003415910992771387, + 0.002066590590402484, + 0.0021131110843271017, + 0.003189347218722105, + 0.0016176776262000203, + 0.0029081853572279215, + 0.0013625107239931822, + 0.0017153591616079211, + 0.0017624126048758626, + 0.0029768813401460648, + 0.0016614451305940747, + 0.002257466781884432, + 0.0021961373277008533, + 0.0020905202254652977, + 0.002667803782969713, + 0.0024761308450251818, + 0.0031739147379994392, + 0.0035764439962804317, + 0.0023443219251930714, + 0.003069814993068576, + 0.001887982594780624, + 0.0037544218357652426, + 0.00540518295019865, + 0.020214764401316643, + 0.004424883984029293, + 0.005409078672528267, + 0.009187760762870312, + 0.017428360879421234, + 0.005279912147670984, + 0.0057085552252829075, + 0.0024626695085316896, + 0.016752589493989944, + 0.001984555972740054, + 0.002786974422633648, + 0.005880697164684534, + 0.017444251105189323, + 0.0046601006761193275, + 0.0019277808023616672, + 0.009456672705709934, + 0.04061315953731537, + 0.006860597524791956, + 0.008586644195020199, + 0.03233906626701355, + 0.032512374222278595, + 0.012622779235243797, + 0.010697593912482262, + 0.019781284034252167, + 0.1827133744955063, + 0.014189125970005989, + 0.014596535824239254, + 0.06722328811883926, + 0.15511879324913025, + 0.013261069543659687, + 0.016739575192332268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.002819731365889311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up from a slightly upward angle held still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a classic dynamic microphone with a spherical grille and cylindrical chrome body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the microphone slowly tilt downward, rub against its plastic clamp, and then break free to fall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the microphone strike the polished wooden floor, bounce, roll, and come to rest at the edge of a spotlight circle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a single vertical cold-white spotlight creating high contrast with deep shadows from a velvet curtain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810613602399826 + ], + "opposite_logprob_tokens": [ + -3.2881062030792236 + ], + "logprob_reward": -0.03810613602399826, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810613602399826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04731297865509987, + 0.0250847227871418, + 0.04934419319033623, + 0.026752332225441933, + 0.008091846480965614, + 0.019923865795135498, + 0.04139907285571098, + 0.04873204231262207, + 0.03161671757698059, + 0.7017422318458557 + ], + "credit_map": [ + 0.004790394101291895, + 0.003969097509980202, + 0.009186855517327785, + 0.004457472357898951, + 0.0030243529472500086, + 0.007882244884967804, + 0.004679800942540169, + 0.009322763420641422, + 0.002430523745715618, + 0.0023905232083052397, + 0.003555063856765628, + 0.003055357374250889, + 0.001555738621391356, + 0.003955190535634756, + 0.002840882632881403, + 0.005301445722579956, + 0.0037792003713548183, + 0.0035864883102476597, + 0.008389029651880264, + 0.011327334679663181, + 0.0036592329852283, + 0.007584906183183193, + 0.005141728557646275, + 0.005876277107745409, + 0.0023057463113218546, + 0.004858090076595545, + 0.004446233622729778, + 0.0034494856372475624, + 0.0023988799657672644, + 0.004244867246598005, + 0.0018328429432585835, + 0.003216188633814454, + 0.0007785942871123552, + 0.0006265544216148555, + 0.000847869785502553, + 0.001383910421282053, + 0.0006667869165539742, + 0.0012983371270820498, + 0.000946778804063797, + 0.0015430155908688903, + 0.002598533174023032, + 0.0019685099832713604, + 0.003329564817249775, + 0.002935154363512993, + 0.0013244538567960262, + 0.002839521039277315, + 0.0017312208656221628, + 0.0031969083938747644, + 0.002581060631200671, + 0.010207148268818855, + 0.002885352587327361, + 0.0035140998661518097, + 0.005707264877855778, + 0.009922781959176064, + 0.0027359328232705593, + 0.0038454330060631037, + 0.0016532365698367357, + 0.01895664632320404, + 0.0015121971955522895, + 0.0015512488316744566, + 0.003092189086601138, + 0.018235838040709496, + 0.0017363415099680424, + 0.001994348829612136, + 0.002266718540340662, + 0.00648862961679697, + 0.00134332908783108, + 0.001966137671843171, + 0.0047310576774179935, + 0.009425219148397446, + 0.0024193532299250364, + 0.0029762759804725647, + 0.016718633472919464, + 0.2860085368156433, + 0.014779304154217243, + 0.013527723029255867, + 0.08231495320796967, + 0.25536128878593445, + 0.010538754984736443, + 0.02249304950237274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.03810613602399826 + } + ], + "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_9/index_tensor([6240, 6240])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_9/index_tensor([6240, 6240])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..48e1893932f0241689db7f3e4fce885b02c771d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_9/index_tensor([6240, 6240])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle close-up captures a clandestine gathering of human-shaped candies huddled together on a polished beige granite kitchen counter. At the center of the group, a translucent red gummy bear, its gelatinous body shimmering, leans forward conspiratorially, its body slightly compressed. Beside it, a sunny yellow gummy bear gestures excitedly with one slender arm, its entire body swaying with silent, urgent movement. A tall, spinning lollipop, its hard candy surface a smooth spiral of pink and white, slightly separated and resting on a white paper strip, its shape tilted toward the bear, as if listening intently. A warm, soft glow from an invisible light beneath the cabinet causes the gummy bear to glow from within, casting sharp, sweet highlights on the lollipop's crystal surface and soft reflections on the counter. In the deeply out-of-focus background, the blurred shapes of a stainless steel toaster and a ceramic flour jar create the sense of a larger, sleepy world, heightening the whimsical secrecy and focused intimacy of the tiny, silent conversation.", + "seed": 23444, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_9/index_6240_seed_23444_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": -0.15683367624878883, + "vlm_logprob": -0.15683367624878883, + "vlm_yes_no_margin": 2.549999988079071, + "vlm_token_credit": -0.15683367624878883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3175528645515442 + ], + "opposite_logprob_tokens": [ + -1.3175528049468994 + ], + "logprob_reward": -0.3175528645515442, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3175528645515442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3438929319381714, + 0.04680543392896652, + 0.04975244775414467, + 0.04970115050673485, + 0.050475846976041794, + 0.07349446415901184, + 0.1403300017118454, + 0.12903349101543427, + 0.0925121083855629, + 0.024002129212021828 + ], + "credit_map": [ + 0.03502234071493149, + 0.05389334261417389, + 0.03669482842087746, + 0.07672319561243057, + 0.03479587286710739, + 0.04052731767296791, + 0.029283633455634117, + 0.0369524210691452, + 0.006100414786487818, + 0.004370904993265867, + 0.004742924124002457, + 0.008828254416584969, + 0.006074927747249603, + 0.007609418127685785, + 0.004455976653844118, + 0.00462261401116848, + 0.004108819179236889, + 0.0076978192664682865, + 0.005978710949420929, + 0.0098651722073555, + 0.0051819272339344025, + 0.005797955207526684, + 0.004714955110102892, + 0.006407089997082949, + 0.004345867782831192, + 0.0073091075755655766, + 0.004273414611816406, + 0.010532868094742298, + 0.005506583955138922, + 0.005746033973991871, + 0.005421485751867294, + 0.00656579015776515, + 0.005397796630859375, + 0.003985840827226639, + 0.004272296093404293, + 0.009081458672881126, + 0.006539671216160059, + 0.007498714607208967, + 0.005835796240717173, + 0.007864272221922874, + 0.006086903624236584, + 0.011737659573554993, + 0.007815058343112469, + 0.015867840498685837, + 0.007196793798357248, + 0.008419342339038849, + 0.0073202974162995815, + 0.009050572291016579, + 0.011799338273704052, + 0.023377271369099617, + 0.013688855804502964, + 0.019777709618210793, + 0.017124561592936516, + 0.01791982911527157, + 0.013854343444108963, + 0.022788096219301224, + 0.014879697933793068, + 0.022625906392931938, + 0.012656845152378082, + 0.01711459644138813, + 0.01566217839717865, + 0.023213602602481842, + 0.0106698889285326, + 0.012210765853524208, + 0.006991823203861713, + 0.0168292298913002, + 0.009019956924021244, + 0.023003319278359413, + 0.008138093166053295, + 0.008615688420832157, + 0.007928837090730667, + 0.011985166929662228, + 0.0022918598260730505, + 0.0037678470835089684, + 0.0025198482908308506, + 0.0055504306219518185, + 0.0024503283202648163, + 0.002428109524771571, + 0.0024444074369966984, + 0.0025492971763014793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.3175528645515442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088193282485008 + ], + "opposite_logprob_tokens": [ + -4.141088008880615 + ], + "logprob_reward": -0.016088193282485008, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088193282485008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2072175145149231, + 0.11476312577724457, + 0.07000809162855148, + 0.06948354095220566, + 0.044697411358356476, + 0.08243657648563385, + 0.26740598678588867, + 0.06650698184967041, + 0.02832847274839878, + 0.04915238171815872 + ], + "credit_map": [ + 0.01647677831351757, + 0.03542109206318855, + 0.021759409457445145, + 0.05117495357990265, + 0.018200188875198364, + 0.019114168360829353, + 0.019463617354631424, + 0.025607304647564888, + 0.009588135406374931, + 0.01896812953054905, + 0.008530175313353539, + 0.02383912168443203, + 0.013618274591863155, + 0.012401730753481388, + 0.010264786891639233, + 0.01755276881158352, + 0.005335085093975067, + 0.007708583027124405, + 0.007545379921793938, + 0.012287222780287266, + 0.007928002625703812, + 0.008376962505280972, + 0.009057902731001377, + 0.011768958531320095, + 0.005929754581302404, + 0.010003541596233845, + 0.006942772306501865, + 0.01222540345042944, + 0.009167204611003399, + 0.009625168517231941, + 0.007082319352775812, + 0.008507371880114079, + 0.004190297797322273, + 0.005619284696877003, + 0.004637282807379961, + 0.007101576309651136, + 0.0060725137591362, + 0.006644503213465214, + 0.004577523097395897, + 0.005854427814483643, + 0.007122286595404148, + 0.011905449442565441, + 0.009350110776722431, + 0.01669052056968212, + 0.005516964476555586, + 0.010033225640654564, + 0.008670072071254253, + 0.013147951103746891, + 0.02164154127240181, + 0.036181237548589706, + 0.02575644478201866, + 0.06323204934597015, + 0.028731169179081917, + 0.028708379715681076, + 0.025957878679037094, + 0.03719726577401161, + 0.00737685477361083, + 0.01092782523483038, + 0.0062001412734389305, + 0.014947349205613136, + 0.0047989170998334885, + 0.006802258547395468, + 0.006368658039718866, + 0.009084980003535748, + 0.0024955084081739187, + 0.003580526215955615, + 0.0029905803967267275, + 0.0030868558678776026, + 0.002758461283519864, + 0.003849346423521638, + 0.0030385886784642935, + 0.006528603844344616, + 0.005291749257594347, + 0.004471268970519304, + 0.0045766159892082214, + 0.008391992188990116, + 0.005740019492805004, + 0.007093297317624092, + 0.004863186273723841, + 0.008724250830709934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.016088193282485008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794966802001 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794966802001, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794966802001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17784219980239868, + 0.07346585392951965, + 0.18440403044223785, + 0.06781274825334549, + 0.04882889240980148, + 0.16812464594841003, + 0.1337585598230362, + 0.079657182097435, + 0.03195429965853691, + 0.03415155038237572 + ], + "credit_map": [ + 0.01568952202796936, + 0.023945970460772514, + 0.018497319892048836, + 0.03644949942827225, + 0.012582316063344479, + 0.02305477298796177, + 0.020496053621172905, + 0.027126751840114594, + 0.0071326326578855515, + 0.013070368207991123, + 0.008166534826159477, + 0.015897469595074654, + 0.005255027674138546, + 0.007368722464889288, + 0.007441958412528038, + 0.009133142419159412, + 0.014207939617335796, + 0.0395207442343235, + 0.018932754173874855, + 0.051906079053878784, + 0.012066520750522614, + 0.018637381494045258, + 0.014934822916984558, + 0.014197787269949913, + 0.007383174728602171, + 0.009645906277000904, + 0.0060835061594843864, + 0.01244931761175394, + 0.006450775545090437, + 0.012191174551844597, + 0.0061975945718586445, + 0.0074113039299845695, + 0.005714549217373133, + 0.007207235787063837, + 0.004942982457578182, + 0.00943504273891449, + 0.0035256571136415005, + 0.00856106448918581, + 0.0050518810749053955, + 0.00439047534018755, + 0.012659456580877304, + 0.030435021966695786, + 0.01710248365998268, + 0.03859078511595726, + 0.015679731965065002, + 0.019147759303450584, + 0.01372537948191166, + 0.020784026011824608, + 0.010380922816693783, + 0.016490111127495766, + 0.013066291809082031, + 0.028383949771523476, + 0.01558706071227789, + 0.015663497149944305, + 0.013058256357908249, + 0.02112847939133644, + 0.007016800343990326, + 0.0101327458396554, + 0.007911702618002892, + 0.013592728413641453, + 0.008980209939181805, + 0.00920200627297163, + 0.008471394889056683, + 0.014349592849612236, + 0.002879190491512418, + 0.005053545348346233, + 0.0031163024250417948, + 0.0068502165377140045, + 0.0029145809821784496, + 0.00321937445551157, + 0.0029936591163277626, + 0.004927428439259529, + 0.004744836129248142, + 0.004106045234948397, + 0.003340945579111576, + 0.006132860668003559, + 0.002827469725161791, + 0.006539556663483381, + 0.003004124853760004, + 0.003455712692812085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.029794966802001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368622809648514 + ], + "opposite_logprob_tokens": [ + -3.4086861610412598 + ], + "logprob_reward": -0.03368622809648514, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368622809648514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2703606188297272, + 0.06508298218250275, + 0.17063133418560028, + 0.06686162948608398, + 0.059493400156497955, + 0.08325517922639847, + 0.13629241287708282, + 0.048833925276994705, + 0.06182288005948067, + 0.03736567124724388 + ], + "credit_map": [ + 0.022275766357779503, + 0.04377209395170212, + 0.02480253577232361, + 0.03957553207874298, + 0.02898711897432804, + 0.034867990761995316, + 0.031010255217552185, + 0.04506931081414223, + 0.008065194822847843, + 0.007853185757994652, + 0.0066823395900428295, + 0.010003853589296341, + 0.008290487341582775, + 0.009164751507341862, + 0.005984350573271513, + 0.009038817137479782, + 0.01905386708676815, + 0.019374050199985504, + 0.014729934744536877, + 0.02201119437813759, + 0.023797737434506416, + 0.03337154537439346, + 0.017754629254341125, + 0.020538372918963432, + 0.006682637147605419, + 0.006329966243356466, + 0.007191247306764126, + 0.009738245978951454, + 0.009974116459488869, + 0.008672792464494705, + 0.008574885316193104, + 0.009697739034891129, + 0.005851656664162874, + 0.005060188937932253, + 0.005782920401543379, + 0.008397368714213371, + 0.010118345730006695, + 0.006888777948915958, + 0.008089353330433369, + 0.009304791688919067, + 0.008165732026100159, + 0.00863781850785017, + 0.007919484749436378, + 0.012582777068018913, + 0.011309093795716763, + 0.014463873580098152, + 0.008364529348909855, + 0.011811874806880951, + 0.011463752016425133, + 0.023893199861049652, + 0.013334488496184349, + 0.037935156375169754, + 0.010223707184195518, + 0.01289219968020916, + 0.013108043000102043, + 0.013441876508295536, + 0.0059812129475176334, + 0.006450539454817772, + 0.004857921972870827, + 0.008575302548706532, + 0.005497322417795658, + 0.007378887385129929, + 0.004604648798704147, + 0.0054880897514522076, + 0.004650089889764786, + 0.010748263448476791, + 0.005358180496841669, + 0.015203351154923439, + 0.0049939583986997604, + 0.005726718343794346, + 0.006578121799975634, + 0.008564195595681667, + 0.003434813814237714, + 0.0038994615897536278, + 0.0038261646404862404, + 0.00682846549898386, + 0.005423638503998518, + 0.005614139139652252, + 0.003818854922428727, + 0.004520133137702942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.03368622809648514 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704612851142883 + ], + "opposite_logprob_tokens": [ + -1.1370460987091064 + ], + "logprob_reward": -0.38704612851142883, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38704612851142883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20154313743114471, + 0.13588552176952362, + 0.08456365764141083, + 0.043816130608320236, + 0.05044254660606384, + 0.07130146771669388, + 0.19748321175575256, + 0.07754341512918472, + 0.07553233206272125, + 0.061888545751571655 + ], + "credit_map": [ + 0.019881343469023705, + 0.03237410634756088, + 0.018306488171219826, + 0.02841266430914402, + 0.031291864812374115, + 0.03218667581677437, + 0.01625826396048069, + 0.022831754758954048, + 0.013606035150587559, + 0.019121911376714706, + 0.015172592364251614, + 0.03182154893875122, + 0.009039885364472866, + 0.013628794811666012, + 0.014833947643637657, + 0.018660807982087135, + 0.007926049642264843, + 0.01259582582861185, + 0.008248785510659218, + 0.014770684763789177, + 0.00803559273481369, + 0.009118865244090557, + 0.009846366010606289, + 0.014021492563188076, + 0.004930740687996149, + 0.005328884813934565, + 0.004549770150333643, + 0.006960451602935791, + 0.005682926159352064, + 0.006033587735146284, + 0.004525316879153252, + 0.005804454442113638, + 0.00524885905906558, + 0.007299236953258514, + 0.005930440966039896, + 0.008108910173177719, + 0.005769162438809872, + 0.007833702489733696, + 0.004697293508797884, + 0.005554946139454842, + 0.0063616130501031876, + 0.009770963340997696, + 0.00840036291629076, + 0.014143763110041618, + 0.005938044749200344, + 0.0060806963592767715, + 0.008194142952561378, + 0.012411880306899548, + 0.017918573692440987, + 0.02779579348862171, + 0.019917543977499008, + 0.05072503164410591, + 0.017638644203543663, + 0.02249418944120407, + 0.01906077191233635, + 0.02193266898393631, + 0.00805772002786398, + 0.008326996117830276, + 0.007446143310517073, + 0.011415315791964531, + 0.0082115288823843, + 0.011621896177530289, + 0.009227142669260502, + 0.013236678205430508, + 0.007017781492322683, + 0.01216469332575798, + 0.007117537781596184, + 0.01381415780633688, + 0.008925956673920155, + 0.012122038751840591, + 0.006837569177150726, + 0.007532599847763777, + 0.00702336709946394, + 0.006329051218926907, + 0.006833017338067293, + 0.011235922574996948, + 0.0051075490191578865, + 0.008649027906358242, + 0.006204949226230383, + 0.010505666956305504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.38704612851142883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a clandestine gathering of human-shaped candies huddled together on a polished beige granite kitchen counter. At the center of the group, a translucent red gummy bear, its gelatinous body shimmering, leans forward conspiratorially, its body slightly compressed. Beside it, a sunny yellow gummy bear gestures excitedly with one slender arm, its entire body swaying with silent, urgent movement. A tall, spinning lollipop, its hard candy surface a smooth spiral of pink and white, slightly separated and resting on a white paper strip, its shape tilted toward the bear, as if listening intently. A warm, soft glow from an invisible light beneath the cabinet causes the gummy bear to glow from within, casting sharp, sweet highlights on the lollipop's crystal surface and soft reflections on the counter. In the deeply out-of-focus background, the blurred shapes of a stainless steel toaster and a ceramic flour jar create the sense of a larger, sleepy world, heightening the whimsical secrecy and focused intimacy of the tiny, silent conversation.", + "seed": 23445, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_9/index_6240_seed_23445_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": -0.1287664184346795, + "vlm_logprob": -0.1287664184346795, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.1287664184346795, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25638794898986816 + ], + "opposite_logprob_tokens": [ + -1.5063879489898682 + ], + "logprob_reward": -0.25638794898986816, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25638794898986816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06284759193658829, + 0.1822386085987091, + 0.07908877730369568, + 0.09309691935777664, + 0.04623597115278244, + 0.2178211808204651, + 0.08572550863027573, + 0.08524589240550995, + 0.11279778182506561, + 0.03490176796913147 + ], + "credit_map": [ + 0.007100522983819246, + 0.008904065936803818, + 0.00520453741773963, + 0.0062186275608837605, + 0.008200941607356071, + 0.00961246807128191, + 0.010139363817870617, + 0.007467065006494522, + 0.0072560617700219154, + 0.004793443717062473, + 0.0072578429244458675, + 0.02081332355737686, + 0.015405463986098766, + 0.02855980023741722, + 0.049349162727594376, + 0.048803508281707764, + 0.006365866865962744, + 0.004947760608047247, + 0.005566224921494722, + 0.024993352591991425, + 0.005756746511906385, + 0.013318521901965141, + 0.009727953933179379, + 0.00841235090047121, + 0.006652493495494127, + 0.00491167651489377, + 0.004558470565825701, + 0.03603227809071541, + 0.0046409727074205875, + 0.008938516490161419, + 0.012072322890162468, + 0.01529018860310316, + 0.003124152310192585, + 0.0034349069464951754, + 0.002943690400570631, + 0.011578678153455257, + 0.004811154678463936, + 0.009673547931015491, + 0.007086065597832203, + 0.0035837721079587936, + 0.009378678165376186, + 0.007471815682947636, + 0.009487894363701344, + 0.047420259565114975, + 0.01968172751367092, + 0.0191891361027956, + 0.06886132806539536, + 0.03633033111691475, + 0.00528025534003973, + 0.0037459235172718763, + 0.005586078390479088, + 0.008820168673992157, + 0.010600268840789795, + 0.009005293250083923, + 0.023473074659705162, + 0.019214443862438202, + 0.004501227755099535, + 0.002673504641279578, + 0.0037668931763619184, + 0.011651504784822464, + 0.009050298482179642, + 0.007006767671555281, + 0.028133178129792213, + 0.018462523818016052, + 0.009260834194719791, + 0.007339135743677616, + 0.010486708022654057, + 0.01490399707108736, + 0.013597661629319191, + 0.0194588303565979, + 0.020879166200757027, + 0.01687144860625267, + 0.004896511789411306, + 0.002714678877964616, + 0.004076594021171331, + 0.0048289126716554165, + 0.004257259424775839, + 0.0067511326633393764, + 0.003958037588745356, + 0.0034186416305601597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.25638794898986816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609441079199314 + ], + "opposite_logprob_tokens": [ + -4.141094207763672 + ], + "logprob_reward": -0.01609441079199314, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609441079199314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02635275572538376, + 0.10622632503509521, + 0.046153146773576736, + 0.030200079083442688, + 0.019999662414193153, + 0.6177719831466675, + 0.036100395023822784, + 0.05414874479174614, + 0.028212741017341614, + 0.03483421355485916 + ], + "credit_map": [ + 0.003112503094598651, + 0.003202836262062192, + 0.0022103418596088886, + 0.002476142253726721, + 0.0035740083549171686, + 0.004871806595474482, + 0.003965039737522602, + 0.0029400759376585484, + 0.006941745989024639, + 0.0032151436898857355, + 0.010104866698384285, + 0.009780069813132286, + 0.005751407705247402, + 0.017884880304336548, + 0.014866449870169163, + 0.037681758403778076, + 0.003858250332996249, + 0.003351779654622078, + 0.0033605555072426796, + 0.014802246354520321, + 0.0033296202309429646, + 0.005083200987428427, + 0.004841251764446497, + 0.007526242174208164, + 0.0025842024479061365, + 0.0019720355048775673, + 0.0037184562534093857, + 0.0027325432747602463, + 0.0022843019105494022, + 0.006448027677834034, + 0.0031114043667912483, + 0.007349106483161449, + 0.0017016874626278877, + 0.0013811668613925576, + 0.002093301387503743, + 0.0022785987239331007, + 0.0020210666116327047, + 0.004960997961461544, + 0.0028020571917295456, + 0.0027607837691903114, + 0.028706369921565056, + 0.013957265764474869, + 0.020390477031469345, + 0.07713637501001358, + 0.05629916116595268, + 0.0681958869099617, + 0.1948278695344925, + 0.1582585573196411, + 0.001990694785490632, + 0.001135095488280058, + 0.0018576054135337472, + 0.0031558896880596876, + 0.00269561936147511, + 0.0054733105935156345, + 0.007236231118440628, + 0.012555943801999092, + 0.0029827007092535496, + 0.0020166097674518824, + 0.0020860638469457626, + 0.0070909783244132996, + 0.005180568899959326, + 0.00435326062142849, + 0.016932323575019836, + 0.013506239280104637, + 0.0021741667296737432, + 0.002206280129030347, + 0.0024673077277839184, + 0.004407975357025862, + 0.0035739510785788298, + 0.003692393423989415, + 0.004553722217679024, + 0.005136942956596613, + 0.0034539292100816965, + 0.0018108206568285823, + 0.003392841899767518, + 0.0030335180927067995, + 0.005805263761430979, + 0.0025837491266429424, + 0.01062784530222416, + 0.00412624329328537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.01609441079199314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107827853411436 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.01107827853411436, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107827853411436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033858541399240494, + 0.3187911808490753, + 0.023972000926733017, + 0.0442245714366436, + 0.02265920303761959, + 0.2730604112148285, + 0.04988246411085129, + 0.17758269608020782, + 0.029460355639457703, + 0.026508543640375137 + ], + "credit_map": [ + 0.003382055088877678, + 0.0033355513587594032, + 0.0030863373540341854, + 0.006243277806788683, + 0.004228522535413504, + 0.0047106659039855, + 0.005151120480149984, + 0.0037210083100944757, + 0.010895541869103909, + 0.008345240727066994, + 0.011421797797083855, + 0.05187022686004639, + 0.026853950694203377, + 0.04468156769871712, + 0.09043650329113007, + 0.07428634911775589, + 0.0021505248732864857, + 0.0036461236886680126, + 0.0019397136056795716, + 0.003652820596471429, + 0.003256875555962324, + 0.003232482820749283, + 0.003423486603423953, + 0.002669972600415349, + 0.0030753978062421083, + 0.0036220813635736704, + 0.003409836208447814, + 0.014864311553537846, + 0.002828311873599887, + 0.0070151337422430515, + 0.0060645295307040215, + 0.0033449712209403515, + 0.0017081622499972582, + 0.0030104368925094604, + 0.0018173264106735587, + 0.004668577574193478, + 0.002489226870238781, + 0.003503198502585292, + 0.0030022042337805033, + 0.002460068091750145, + 0.012838506139814854, + 0.012396136298775673, + 0.017636368051171303, + 0.08677011728286743, + 0.025372520089149475, + 0.008327190764248371, + 0.07453233748674393, + 0.035187240689992905, + 0.0034254207275807858, + 0.0037713018245995045, + 0.0034471030812710524, + 0.012272167950868607, + 0.004451802931725979, + 0.00472658034414053, + 0.004719228483736515, + 0.013068854808807373, + 0.011101148091256618, + 0.0052006240002810955, + 0.01158942561596632, + 0.03268010914325714, + 0.02052806131541729, + 0.009652302600443363, + 0.06205171346664429, + 0.02477930672466755, + 0.0027556121349334717, + 0.003426157869398594, + 0.002505670301616192, + 0.003898019203916192, + 0.005328836850821972, + 0.002802702598273754, + 0.0045268055982887745, + 0.004216549452394247, + 0.0024182917550206184, + 0.002233607927337289, + 0.002882551634684205, + 0.0031240663956850767, + 0.0045782518573105335, + 0.0024903223384171724, + 0.006123277824372053, + 0.0026581736747175455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.01107827853411436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897159457206726 + ], + "opposite_logprob_tokens": [ + -2.5789716243743896 + ], + "logprob_reward": -0.07897159457206726, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897159457206726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0437026284635067, + 0.2535165250301361, + 0.04119354858994484, + 0.022962581366300583, + 0.02250685915350914, + 0.11602182686328888, + 0.35652950406074524, + 0.07581828534603119, + 0.041254863142967224, + 0.02649340033531189 + ], + "credit_map": [ + 0.004593285731971264, + 0.004031366668641567, + 0.004162067547440529, + 0.0037798385601490736, + 0.006504420191049576, + 0.009299512021243572, + 0.005390687845647335, + 0.0059414515271782875, + 0.011054822243750095, + 0.004198599141091108, + 0.008354139514267445, + 0.013102319091558456, + 0.020009709522128105, + 0.04127495363354683, + 0.06786647439002991, + 0.08765552937984467, + 0.0034617255441844463, + 0.003117828629910946, + 0.003073338884860277, + 0.006266740150749683, + 0.005244637839496136, + 0.008567376993596554, + 0.006020638160407543, + 0.005441260989755392, + 0.0018911539809778333, + 0.0017627516062930226, + 0.0013508162228390574, + 0.006865420378744602, + 0.001517835888080299, + 0.0029070936143398285, + 0.0026350237894803286, + 0.004032486584037542, + 0.0016558722127228975, + 0.0030968375504016876, + 0.0015103926416486502, + 0.004878639243543148, + 0.0025060949847102165, + 0.0035562664270401, + 0.003236963879317045, + 0.002065792679786682, + 0.005394105799496174, + 0.005594540853053331, + 0.005006776656955481, + 0.03957625851035118, + 0.011225178837776184, + 0.0034723931457847357, + 0.035196661949157715, + 0.010555910877883434, + 0.017485806718468666, + 0.00715941097587347, + 0.014783837832510471, + 0.029241062700748444, + 0.03832181915640831, + 0.03206205368041992, + 0.11896278709173203, + 0.09851271659135818, + 0.005357158370316029, + 0.0032436037436127663, + 0.006925180554389954, + 0.014128309674561024, + 0.009025304578244686, + 0.005027637351304293, + 0.02475605346262455, + 0.007355034351348877, + 0.003446429967880249, + 0.003959524445235729, + 0.004390954039990902, + 0.006574386730790138, + 0.0060344357043504715, + 0.005827577784657478, + 0.004382286220788956, + 0.0066392673179507256, + 0.0021606346126645803, + 0.0027700881473720074, + 0.0034734054934233427, + 0.0030999677255749702, + 0.0038275059778243303, + 0.0032236366532742977, + 0.005100716836750507, + 0.00283744465559721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.07897159457206726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812998592853546 + ], + "opposite_logprob_tokens": [ + -1.4062998294830322 + ], + "logprob_reward": -0.2812998592853546, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812998592853546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046128321439027786, + 0.24166494607925415, + 0.02611607126891613, + 0.04536573588848114, + 0.06182239577174187, + 0.053876329213380814, + 0.05734242871403694, + 0.41852861642837524, + 0.02817576378583908, + 0.020979389548301697 + ], + "credit_map": [ + 0.0049375006929039955, + 0.0052556549198925495, + 0.003223380772396922, + 0.011688092723488808, + 0.003871932625770569, + 0.006565017160028219, + 0.005112848710268736, + 0.0054738945327699184, + 0.01054657343775034, + 0.006399016827344894, + 0.010417100973427296, + 0.03871811553835869, + 0.02167128585278988, + 0.03398488461971283, + 0.06851664930582047, + 0.05141132324934006, + 0.001902761054225266, + 0.0018414255464449525, + 0.0013848530361428857, + 0.008957010693848133, + 0.0018843174912035465, + 0.0033515342511236668, + 0.003291437868028879, + 0.0035027312114834785, + 0.003042489755898714, + 0.003257616190239787, + 0.0024998565204441547, + 0.016111787408590317, + 0.003179380903020501, + 0.00593002000823617, + 0.006386549212038517, + 0.004958040080964565, + 0.004667071159929037, + 0.0035026301629841328, + 0.002914269221946597, + 0.020865054801106453, + 0.0046771736815571785, + 0.010104440152645111, + 0.010360240936279297, + 0.004731511231511831, + 0.0037173426244407892, + 0.00264610699377954, + 0.0049499706365168095, + 0.004832415841519833, + 0.005548408720642328, + 0.008359693922102451, + 0.011655072681605816, + 0.012167316861450672, + 0.003331821644678712, + 0.002719099633395672, + 0.0030218425672501326, + 0.00658019445836544, + 0.004752959590405226, + 0.008435952477157116, + 0.010183734819293022, + 0.01831682398915291, + 0.025446860119700432, + 0.016578197479248047, + 0.029400641098618507, + 0.07927203923463821, + 0.05047982558608055, + 0.025132423266768456, + 0.14067474007606506, + 0.051543883979320526, + 0.0023685311898589134, + 0.001936278771609068, + 0.0031597164925187826, + 0.004055151715874672, + 0.0036060272250324488, + 0.004726265091449022, + 0.004251818172633648, + 0.004071973264217377, + 0.0021164051722735167, + 0.0009594057337380946, + 0.002030831528827548, + 0.0019573457539081573, + 0.0028545598033815622, + 0.002961731981486082, + 0.006354213692247868, + 0.0017448966391384602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.2812998592853546 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a clandestine gathering of human-shaped candies huddled together on a polished beige granite kitchen counter. At the center of the group, a translucent red gummy bear, its gelatinous body shimmering, leans forward conspiratorially, its body slightly compressed. Beside it, a sunny yellow gummy bear gestures excitedly with one slender arm, its entire body swaying with silent, urgent movement. A tall, spinning lollipop, its hard candy surface a smooth spiral of pink and white, slightly separated and resting on a white paper strip, its shape tilted toward the bear, as if listening intently. A warm, soft glow from an invisible light beneath the cabinet causes the gummy bear to glow from within, casting sharp, sweet highlights on the lollipop's crystal surface and soft reflections on the counter. In the deeply out-of-focus background, the blurred shapes of a stainless steel toaster and a ceramic flour jar create the sense of a larger, sleepy world, heightening the whimsical secrecy and focused intimacy of the tiny, silent conversation.", + "seed": 23446, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_9/index_6240_seed_23446_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": -0.22028433308005332, + "vlm_logprob": -0.22028433308005332, + "vlm_yes_no_margin": 2.050000011920929, + "vlm_token_credit": -0.22028433308005332, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5303078293800354 + ], + "opposite_logprob_tokens": [ + -0.9053078293800354 + ], + "logprob_reward": -0.5303078293800354, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5303078293800354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14823086559772491, + 0.3661685287952423, + 0.12573330104351044, + 0.04514120891690254, + 0.08270172774791718, + 0.05885768309235573, + 0.09328527003526688, + 0.018066545948386192, + 0.027142172679305077, + 0.03467271476984024 + ], + "credit_map": [ + 0.024291744455695152, + 0.029130514711141586, + 0.021742044016718864, + 0.014045679941773415, + 0.0169020127505064, + 0.011661972850561142, + 0.010759366676211357, + 0.01969752088189125, + 0.0291191004216671, + 0.049707163125276566, + 0.10564146935939789, + 0.027036216109991074, + 0.08642710000276566, + 0.024541979655623436, + 0.022600578144192696, + 0.021094920113682747, + 0.021332956850528717, + 0.027605203911662102, + 0.025667933747172356, + 0.010517440736293793, + 0.011877457611262798, + 0.007146063260734081, + 0.0075876121409237385, + 0.013998634181916714, + 0.004872054327279329, + 0.007725446950644255, + 0.0053871707059443, + 0.0063711004331707954, + 0.005475841462612152, + 0.005718565080314875, + 0.00478358194231987, + 0.004807447083294392, + 0.011428029276430607, + 0.020264947786927223, + 0.01052518654614687, + 0.008898558095097542, + 0.005622801370918751, + 0.010045824572443962, + 0.0066909645684063435, + 0.009225415997207165, + 0.007553286850452423, + 0.013072528876364231, + 0.006716886069625616, + 0.00848141498863697, + 0.005295180715620518, + 0.0053202910348773, + 0.006664834916591644, + 0.005753261502832174, + 0.008609087206423283, + 0.013581902720034122, + 0.009050266817212105, + 0.011457715183496475, + 0.008246701210737228, + 0.01723056100308895, + 0.012194464914500713, + 0.012914568185806274, + 0.001957865199074149, + 0.0017388510750606656, + 0.002941794926300645, + 0.002592436270788312, + 0.001378206885419786, + 0.0017384246457368135, + 0.002677154028788209, + 0.003041813848540187, + 0.003389779943972826, + 0.00441135186702013, + 0.0024478130508214235, + 0.0040896134451031685, + 0.00310233561322093, + 0.003236938500776887, + 0.0038387577515095472, + 0.002625581342726946, + 0.0035411363933235407, + 0.005383658688515425, + 0.007492650765925646, + 0.0032894008327275515, + 0.003127169096842408, + 0.004128057509660721, + 0.0028279635589569807, + 0.004882678855210543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.5303078293800354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023315560072660446 + ], + "opposite_logprob_tokens": [ + -3.773315668106079 + ], + "logprob_reward": -0.023315560072660446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023315560072660446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2871614396572113, + 0.2535119354724884, + 0.05287604033946991, + 0.1433221697807312, + 0.043922215700149536, + 0.08460880815982819, + 0.035011958330869675, + 0.0319325290620327, + 0.04745443910360336, + 0.020198451355099678 + ], + "credit_map": [ + 0.037983547896146774, + 0.03171395882964134, + 0.053135164082050323, + 0.021295320242643356, + 0.05300897732377052, + 0.02788649871945381, + 0.031425975263118744, + 0.03071201592683792, + 0.019968492910265923, + 0.027127495035529137, + 0.0668954849243164, + 0.02042817510664463, + 0.06672574579715729, + 0.018194442614912987, + 0.017643649131059647, + 0.016528479754924774, + 0.009539439342916012, + 0.012992911972105503, + 0.005691996309906244, + 0.004388578236103058, + 0.0046562994830310345, + 0.00442194240167737, + 0.005394450854510069, + 0.0057904282584786415, + 0.015934934839606285, + 0.027052471414208412, + 0.029361655935645103, + 0.013608723878860474, + 0.01909402199089527, + 0.00953123439103365, + 0.013345734216272831, + 0.015393400564789772, + 0.004940493032336235, + 0.0053762635216116905, + 0.006441629491746426, + 0.0047028642147779465, + 0.0034288805909454823, + 0.005741364322602749, + 0.005746621638536453, + 0.007544103544205427, + 0.01341431774199009, + 0.021559610962867737, + 0.008921926841139793, + 0.010034735314548016, + 0.005628637969493866, + 0.005517780315130949, + 0.010807833634316921, + 0.008723977021872997, + 0.0038137759547680616, + 0.004068231675773859, + 0.005635841749608517, + 0.005022828467190266, + 0.002366318367421627, + 0.0033992333337664604, + 0.006201033014804125, + 0.0045046997256577015, + 0.0041358876042068005, + 0.0065712956711649895, + 0.004443541634827852, + 0.004759286995977163, + 0.0020053561311215162, + 0.0021996533032506704, + 0.005098492372781038, + 0.0027190197724848986, + 0.004625644069164991, + 0.010674944147467613, + 0.005286413244903088, + 0.008222926408052444, + 0.004288330674171448, + 0.0039106206968426704, + 0.004729973152279854, + 0.005715589504688978, + 0.0021732498425990343, + 0.003303602570667863, + 0.0019834262784570456, + 0.002790930215269327, + 0.0013797006104141474, + 0.0020163627341389656, + 0.00398173276335001, + 0.0025694489013403654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.023315560072660446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2134241759777069, + 0.2763039469718933, + 0.051398031413555145, + 0.1282987743616104, + 0.045908473432064056, + 0.06657548248767853, + 0.030136728659272194, + 0.014563127420842648, + 0.13876599073410034, + 0.03462526947259903 + ], + "credit_map": [ + 0.03173600137233734, + 0.030797084793448448, + 0.050408218055963516, + 0.01758842170238495, + 0.017863448709249496, + 0.01713424362242222, + 0.024183016270399094, + 0.023713748902082443, + 0.02372264675796032, + 0.027439214289188385, + 0.0903458520770073, + 0.02084086835384369, + 0.04625411331653595, + 0.018901143223047256, + 0.018912188708782196, + 0.029887914657592773, + 0.005690822843462229, + 0.009069406427443027, + 0.009702347218990326, + 0.006384587846696377, + 0.005707891657948494, + 0.004396362695842981, + 0.00609373115003109, + 0.004352880641818047, + 0.017745306715369225, + 0.03522862493991852, + 0.019590796902775764, + 0.013859623111784458, + 0.015937775373458862, + 0.00794511754065752, + 0.008005080744624138, + 0.009986446239054203, + 0.005193249322474003, + 0.00526187801733613, + 0.00969211757183075, + 0.00629036920145154, + 0.0034310685005038977, + 0.004061688203364611, + 0.007192565593868494, + 0.004785532597452402, + 0.007653305772691965, + 0.011088615283370018, + 0.01274553220719099, + 0.008832983672618866, + 0.006821024231612682, + 0.006328130606561899, + 0.007040693890303373, + 0.00606519402936101, + 0.004109891597181559, + 0.00697092991322279, + 0.0036605154164135456, + 0.004228893667459488, + 0.0026039136573672295, + 0.0019940717611461878, + 0.004473106469959021, + 0.0020954073406755924, + 0.0018173278076574206, + 0.002007303759455681, + 0.0014417700003832579, + 0.0018141668988391757, + 0.0011112205684185028, + 0.0013380298623815179, + 0.0032719075679779053, + 0.0017614003736525774, + 0.017213337123394012, + 0.03917819634079933, + 0.015283015556633472, + 0.020423464477062225, + 0.011441405862569809, + 0.009668564423918724, + 0.012494778260588646, + 0.013063237071037292, + 0.004009576514363289, + 0.004175438079982996, + 0.006013220641762018, + 0.004073311574757099, + 0.0034764830488711596, + 0.004589642398059368, + 0.004532837308943272, + 0.003754757111892104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "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\": \"Are human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270446926355362 + ], + "opposite_logprob_tokens": [ + -2.127044677734375 + ], + "logprob_reward": -0.1270446926355362, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270446926355362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3301263749599457, + 0.29425278306007385, + 0.03126595541834831, + 0.05946328118443489, + 0.06450076401233673, + 0.05277736112475395, + 0.07821686565876007, + 0.012238943949341774, + 0.046348411589860916, + 0.030809225514531136 + ], + "credit_map": [ + 0.03062339685857296, + 0.0210934579372406, + 0.10237772017717361, + 0.025846505537629128, + 0.055716946721076965, + 0.026345450431108475, + 0.036670833826065063, + 0.031452059745788574, + 0.026948971673846245, + 0.04179476574063301, + 0.08070633560419083, + 0.024523258209228516, + 0.03616334870457649, + 0.026045028120279312, + 0.025181883946061134, + 0.032889217138290405, + 0.004559733904898167, + 0.007490601856261492, + 0.003543906845152378, + 0.003251239424571395, + 0.0026766962837427855, + 0.0029640209395438433, + 0.00337648531422019, + 0.0034032727126032114, + 0.005783506203442812, + 0.006243967451155186, + 0.01591615378856659, + 0.005034515634179115, + 0.007416530512273312, + 0.005140233319252729, + 0.007310438435524702, + 0.006617935374379158, + 0.009354625828564167, + 0.019047211855649948, + 0.006957097910344601, + 0.008879149332642555, + 0.004248759709298611, + 0.0036347717978060246, + 0.0068036396987736225, + 0.0055755083449184895, + 0.005986255593597889, + 0.01072670891880989, + 0.007325876969844103, + 0.006875208113342524, + 0.004895377904176712, + 0.004192858003079891, + 0.007671150844544172, + 0.005103928968310356, + 0.010905971750617027, + 0.011308508925139904, + 0.015494820661842823, + 0.005194330587983131, + 0.005937590729445219, + 0.010545858182013035, + 0.006328940857201815, + 0.012500844895839691, + 0.0013557379133999348, + 0.0012570829130709171, + 0.0028914993163198233, + 0.0013228022726252675, + 0.0009759415406733751, + 0.0010681997518986464, + 0.0020911439787596464, + 0.001276535214856267, + 0.004823020193725824, + 0.009701025672256947, + 0.0039297244511544704, + 0.006358643528074026, + 0.003732220735400915, + 0.00507017457857728, + 0.005043304059654474, + 0.007690296042710543, + 0.003175632795318961, + 0.005082076881080866, + 0.0035035228356719017, + 0.004562849644571543, + 0.0021356267388910055, + 0.0036320143844932318, + 0.0053048161789774895, + 0.0034126853570342064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.1270446926355362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704580068588257 + ], + "opposite_logprob_tokens": [ + -1.1370458602905273 + ], + "logprob_reward": -0.38704580068588257, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38704580068588257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18799659609794617, + 0.39314258098602295, + 0.08637192100286484, + 0.054483458399772644, + 0.04342621564865112, + 0.044436171650886536, + 0.09737878292798996, + 0.03247404098510742, + 0.03055717796087265, + 0.02973305992782116 + ], + "credit_map": [ + 0.021561767905950546, + 0.021553006023168564, + 0.046935033053159714, + 0.01704665832221508, + 0.019261887297034264, + 0.014982474036514759, + 0.02186087891459465, + 0.02479490078985691, + 0.031114928424358368, + 0.03690333664417267, + 0.13547110557556152, + 0.025049887597560883, + 0.08645620942115784, + 0.022269871085882187, + 0.022680548951029778, + 0.03319665789604187, + 0.013712740503251553, + 0.023959200829267502, + 0.007342893164604902, + 0.010247751139104366, + 0.00926441140472889, + 0.007599969860166311, + 0.006455449853092432, + 0.007789513096213341, + 0.007120456080883741, + 0.005475394893437624, + 0.014424390159547329, + 0.005087205674499273, + 0.006433089263737202, + 0.003975586500018835, + 0.005455248523503542, + 0.006512091960757971, + 0.006051909644156694, + 0.007301609497517347, + 0.006239290814846754, + 0.004924610257148743, + 0.005930704530328512, + 0.003369940211996436, + 0.004351040814071894, + 0.005257111974060535, + 0.0043401834554970264, + 0.00454765185713768, + 0.009851495735347271, + 0.005099361762404442, + 0.0055579510517418385, + 0.004594328813254833, + 0.004922831431031227, + 0.005522371735423803, + 0.015263729728758335, + 0.024850638583302498, + 0.01124582253396511, + 0.008263364434242249, + 0.008767322637140751, + 0.010996049270033836, + 0.006652388721704483, + 0.01133946143090725, + 0.0040224334225058556, + 0.0033489230554550886, + 0.003474839497357607, + 0.004187803249806166, + 0.002593372715637088, + 0.0036809013690799475, + 0.005935557186603546, + 0.005230212118476629, + 0.004281108733266592, + 0.005875191651284695, + 0.003253537928685546, + 0.004585884977132082, + 0.0029835538007318974, + 0.0031142940279096365, + 0.0029900302179157734, + 0.0034735770896077156, + 0.00418860325589776, + 0.005632521118968725, + 0.0030428553000092506, + 0.0043740831315517426, + 0.0026963618583977222, + 0.002620636485517025, + 0.004139141645282507, + 0.0030388568993657827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.38704580068588257 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a clandestine gathering of human-shaped candies huddled together on a polished beige granite kitchen counter. At the center of the group, a translucent red gummy bear, its gelatinous body shimmering, leans forward conspiratorially, its body slightly compressed. Beside it, a sunny yellow gummy bear gestures excitedly with one slender arm, its entire body swaying with silent, urgent movement. A tall, spinning lollipop, its hard candy surface a smooth spiral of pink and white, slightly separated and resting on a white paper strip, its shape tilted toward the bear, as if listening intently. A warm, soft glow from an invisible light beneath the cabinet causes the gummy bear to glow from within, casting sharp, sweet highlights on the lollipop's crystal surface and soft reflections on the counter. In the deeply out-of-focus background, the blurred shapes of a stainless steel toaster and a ceramic flour jar create the sense of a larger, sleepy world, heightening the whimsical secrecy and focused intimacy of the tiny, silent conversation.", + "seed": 23447, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000084/group_9/index_6240_seed_23447_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": -0.1452388573437929, + "vlm_logprob": -0.1452388573437929, + "vlm_yes_no_margin": 2.674999976158142, + "vlm_token_credit": -0.1452388573437929, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15118467807769775 + ], + "opposite_logprob_tokens": [ + -2.026184558868408 + ], + "logprob_reward": -0.15118467807769775, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15118467807769775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22914820909500122, + 0.10627920925617218, + 0.05960688367486, + 0.06999977678060532, + 0.08128301054239273, + 0.0695168524980545, + 0.12043969333171844, + 0.14389009773731232, + 0.0870942696928978, + 0.03274204209446907 + ], + "credit_map": [ + 0.02421194687485695, + 0.03550548851490021, + 0.040251947939395905, + 0.038149815052747726, + 0.02371278405189514, + 0.01761283352971077, + 0.025382092222571373, + 0.024321304634213448, + 0.013907415792346, + 0.01199192926287651, + 0.01668611355125904, + 0.014900131151080132, + 0.012095920741558075, + 0.009923548437654972, + 0.012150164693593979, + 0.014623992145061493, + 0.006810067221522331, + 0.00970869418233633, + 0.008502046577632427, + 0.010378953069448471, + 0.0067592221312224865, + 0.004073398653417826, + 0.006448118016123772, + 0.006926381029188633, + 0.008318794891238213, + 0.01294459868222475, + 0.011339041404426098, + 0.012039726600050926, + 0.005971507169306278, + 0.005116947460919619, + 0.006398538593202829, + 0.007870621047914028, + 0.009254655800759792, + 0.013675114139914513, + 0.01152355782687664, + 0.013371758162975311, + 0.012621290050446987, + 0.005782154854387045, + 0.008805578574538231, + 0.006248898804187775, + 0.007767846342176199, + 0.012659000232815742, + 0.011019140481948853, + 0.010928835719823837, + 0.007315346505492926, + 0.005926421377807856, + 0.006732643116265535, + 0.007167616859078407, + 0.01300481054931879, + 0.017288682982325554, + 0.015958840027451515, + 0.017924267798662186, + 0.012554011307656765, + 0.009157066233456135, + 0.016495458781719208, + 0.01805654726922512, + 0.01893274299800396, + 0.026737667620182037, + 0.021606706082820892, + 0.018015727400779724, + 0.01718984544277191, + 0.011118730530142784, + 0.013531692326068878, + 0.016756979748606682, + 0.009677543304860592, + 0.01858109049499035, + 0.009217746555805206, + 0.012797662988305092, + 0.015085778199136257, + 0.005594927817583084, + 0.0068165622651577, + 0.009322959929704666, + 0.003186892718076706, + 0.005945824086666107, + 0.003918918780982494, + 0.006023777183145285, + 0.005098400171846151, + 0.0018041949952021241, + 0.00385996769182384, + 0.00290406821295619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.15118467807769775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255878061056137 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.01255878061056137, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255878061056137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3350711762905121, + 0.12832021713256836, + 0.061367399990558624, + 0.05055852234363556, + 0.0560138002038002, + 0.06893300265073776, + 0.0966414287686348, + 0.11958881467580795, + 0.042302556335926056, + 0.041203103959560394 + ], + "credit_map": [ + 0.03339598700404167, + 0.0375257283449173, + 0.06573836505413055, + 0.07847171276807785, + 0.03109564073383808, + 0.017291728407144547, + 0.03738578408956528, + 0.03416621685028076, + 0.014712893404066563, + 0.026514150202274323, + 0.01323213241994381, + 0.01941164955496788, + 0.017602285370230675, + 0.007584966253489256, + 0.016863171011209488, + 0.012398947030305862, + 0.005943972617387772, + 0.010601856745779514, + 0.008923163637518883, + 0.00942368432879448, + 0.007646230049431324, + 0.003993958700448275, + 0.0075335330329835415, + 0.007301001343876123, + 0.004980586003512144, + 0.007909202948212624, + 0.0067075821571052074, + 0.005556599237024784, + 0.005815474316477776, + 0.003663297975435853, + 0.00979172345250845, + 0.006134054623544216, + 0.006076537538319826, + 0.007397377863526344, + 0.0077370950020849705, + 0.009553607553243637, + 0.006627985276281834, + 0.004674906842410564, + 0.006242601666599512, + 0.00770367868244648, + 0.00929463654756546, + 0.00772494450211525, + 0.0155633594840765, + 0.009780241176486015, + 0.007996735163033009, + 0.006150039378553629, + 0.007223992142826319, + 0.005199042148888111, + 0.00991963129490614, + 0.020130446180701256, + 0.01498529314994812, + 0.01031523197889328, + 0.011283526197075844, + 0.0060243308544158936, + 0.011755469255149364, + 0.012227491475641727, + 0.014394085854291916, + 0.01609395444393158, + 0.02204895205795765, + 0.02044755034148693, + 0.010887758806347847, + 0.006576641462743282, + 0.01241240929812193, + 0.016727451235055923, + 0.004930349066853523, + 0.006093887612223625, + 0.004094444215297699, + 0.005981858819723129, + 0.006931292824447155, + 0.00524902855977416, + 0.004704630468040705, + 0.004317058715969324, + 0.0046033719554543495, + 0.0056722089648246765, + 0.0044133649207651615, + 0.007279238197952509, + 0.005548845045268536, + 0.0026863107923418283, + 0.006494475528597832, + 0.004505284130573273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01255878061056137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185080960392952 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185080960392952, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185080960392952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28807252645492554, + 0.22101211547851562, + 0.0624614916741848, + 0.08918988704681396, + 0.05456947535276413, + 0.08420256525278091, + 0.053384456783533096, + 0.0784917026758194, + 0.0451945997774601, + 0.023421213030815125 + ], + "credit_map": [ + 0.02830878086388111, + 0.03842339292168617, + 0.04277971759438515, + 0.05192790552973747, + 0.035768281668424606, + 0.023903241381049156, + 0.03322308883070946, + 0.03373809903860092, + 0.020650256425142288, + 0.055536143481731415, + 0.02295820787549019, + 0.029959239065647125, + 0.03786568343639374, + 0.014427033253014088, + 0.020917238667607307, + 0.0186983160674572, + 0.007366228848695755, + 0.012592272832989693, + 0.006644783541560173, + 0.0071122231893241405, + 0.009023133665323257, + 0.004641434643417597, + 0.009771207347512245, + 0.005310208536684513, + 0.009668348357081413, + 0.016718413680791855, + 0.009695561602711678, + 0.013147088699042797, + 0.012996863573789597, + 0.007363616488873959, + 0.01154828816652298, + 0.008051704615354538, + 0.0069572883658111095, + 0.008782193064689636, + 0.007188665680587292, + 0.009273037314414978, + 0.005899582523852587, + 0.004817422479391098, + 0.006619906984269619, + 0.005031378474086523, + 0.01968308351933956, + 0.007514975965023041, + 0.0191918034106493, + 0.007382531184703112, + 0.008381734602153301, + 0.008554458618164062, + 0.007066252175718546, + 0.00642772763967514, + 0.005195209290832281, + 0.007391342427581549, + 0.00867120735347271, + 0.006594681181013584, + 0.005336986854672432, + 0.004627890884876251, + 0.006664464250206947, + 0.008902676403522491, + 0.010571915656328201, + 0.012039925903081894, + 0.011114473454654217, + 0.013553092256188393, + 0.007263324223458767, + 0.005839971825480461, + 0.008592638187110424, + 0.009516365826129913, + 0.004230824764817953, + 0.008970728144049644, + 0.00534781301394105, + 0.005956711713224649, + 0.007393161300569773, + 0.003276197239756584, + 0.0053672194480896, + 0.004651941359043121, + 0.0028661692049354315, + 0.0031069645192474127, + 0.002800180809572339, + 0.0050978586077690125, + 0.002461811061948538, + 0.0015219812048599124, + 0.003106849966570735, + 0.0024593963753432035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.018185080960392952 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001125812530518 + ], + "opposite_logprob_tokens": [ + -2.6950111389160156 + ], + "logprob_reward": -0.07001125812530518, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001125812530518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29054024815559387, + 0.10542836785316467, + 0.08523837476968765, + 0.1065148338675499, + 0.07722284644842148, + 0.058257922530174255, + 0.05651629716157913, + 0.09837642312049866, + 0.07576734572649002, + 0.04613734781742096 + ], + "credit_map": [ + 0.03118050843477249, + 0.03493376821279526, + 0.0608782060444355, + 0.049831848591566086, + 0.024100646376609802, + 0.01750284433364868, + 0.028642522171139717, + 0.04346990957856178, + 0.010685104876756668, + 0.01305343210697174, + 0.012052301317453384, + 0.013581611216068268, + 0.01791747473180294, + 0.013920491561293602, + 0.013262475840747356, + 0.010955476202070713, + 0.009615362621843815, + 0.01319858431816101, + 0.010241137817502022, + 0.010527709499001503, + 0.0129929780960083, + 0.011215540580451488, + 0.00947638787329197, + 0.007970673032104969, + 0.01326436921954155, + 0.017969951033592224, + 0.012770988047122955, + 0.016976509243249893, + 0.010127831250429153, + 0.010732831433415413, + 0.011952170170843601, + 0.012720185332000256, + 0.008170558139681816, + 0.014337161555886269, + 0.007805094588547945, + 0.00842103362083435, + 0.014016452245414257, + 0.009064764715731144, + 0.009382002055644989, + 0.006025781389325857, + 0.009928535670042038, + 0.007586340419948101, + 0.011781382374465466, + 0.00811015535145998, + 0.005992288701236248, + 0.003553892718628049, + 0.006713260430842638, + 0.0045920670963823795, + 0.005894227419048548, + 0.011411353014409542, + 0.008223284967243671, + 0.006399615202099085, + 0.008341622538864613, + 0.004302673973143101, + 0.00514543242752552, + 0.00679808808490634, + 0.012244466692209244, + 0.011048592627048492, + 0.017721474170684814, + 0.009499862790107727, + 0.014642362482845783, + 0.009663943201303482, + 0.012432379648089409, + 0.011123341508209705, + 0.007471021264791489, + 0.0154709592461586, + 0.009359342977404594, + 0.008603110909461975, + 0.01228918507695198, + 0.005577783565968275, + 0.009634243324398994, + 0.007361703552305698, + 0.007908760569989681, + 0.0056810928508639336, + 0.005096429027616978, + 0.006485224701464176, + 0.006229470018297434, + 0.004594906233251095, + 0.005748888943344355, + 0.004392576403915882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07001125812530518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 human-shaped candies huddled together on a polished beige granite kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a translucent red gummy bear at the center leaning forward with a slightly compressed body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sunny yellow gummy bear beside the red one, gesturing with one arm and swaying?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tall spinning lollipop with a pink and white spiral surface resting on a white paper strip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a warm soft glow from beneath the cabinet causing the gummy bear to glow from within?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742544889450073 + ], + "opposite_logprob_tokens": [ + -0.9742544889450073 + ], + "logprob_reward": -0.4742544889450073, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742544889450073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4000544548034668, + 0.14715105295181274, + 0.06039656698703766, + 0.046530019491910934, + 0.03050921857357025, + 0.05629151314496994, + 0.06007538363337517, + 0.094988152384758, + 0.04716193303465843, + 0.05684170126914978 + ], + "credit_map": [ + 0.05198361352086067, + 0.05357703939080238, + 0.05131649598479271, + 0.07676717638969421, + 0.04118673503398895, + 0.03395742177963257, + 0.05115329101681709, + 0.04011271893978119, + 0.014794752933084965, + 0.018557919189333916, + 0.015190983191132545, + 0.02217746712267399, + 0.022838139906525612, + 0.013479748740792274, + 0.02424357645213604, + 0.01586846634745598, + 0.0051957787945866585, + 0.006705164909362793, + 0.00797950942069292, + 0.008617517538368702, + 0.009195404127240181, + 0.006769714877009392, + 0.008650694973766804, + 0.007282792124897242, + 0.005484032444655895, + 0.0070144799537956715, + 0.00809129886329174, + 0.006095981225371361, + 0.005340284202247858, + 0.004248138517141342, + 0.004238345194607973, + 0.00601746654137969, + 0.0031235110945999622, + 0.004123324993997812, + 0.004907742142677307, + 0.005593688227236271, + 0.0031336375977844, + 0.00224732025526464, + 0.004689076449722052, + 0.0026909224689006805, + 0.013022967614233494, + 0.006631865166127682, + 0.01370992325246334, + 0.005813821218907833, + 0.003928810823708773, + 0.0033546974882483482, + 0.005893975030630827, + 0.003935461863875389, + 0.008178061805665493, + 0.007779004517942667, + 0.007997623644769192, + 0.005893369670957327, + 0.0064508649520576, + 0.006279789377003908, + 0.010004119016230106, + 0.007492562290281057, + 0.011050292290747166, + 0.011425774544477463, + 0.0181611105799675, + 0.009814190678298473, + 0.008950436487793922, + 0.007055528927594423, + 0.014007560908794403, + 0.014523263089358807, + 0.006329352501779795, + 0.0074079567566514015, + 0.005686825141310692, + 0.005802746396511793, + 0.005909958388656378, + 0.003743999172002077, + 0.00576763553544879, + 0.006513465661555529, + 0.005089515820145607, + 0.00871509499847889, + 0.006174528505653143, + 0.009932424873113632, + 0.0070421528071165085, + 0.005826111882925034, + 0.008637674152851105, + 0.005424198228865862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.4742544889450073 + } + ], + "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/0000110/group_0/index_tensor([1911, 1911])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_0/index_tensor([1911, 1911])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..437142e8a0db3aace48a431fdb571a812f49059a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_0/index_tensor([1911, 1911])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, horizontal tracking shot, filmed from a slightly upward angle, follows an exquisitely crafted ancient sedan chair as it slowly levitates through the depths of a mist-shrouded bamboo forest. The sedan is meticulously carved from dark lacquered wood, its frame draped in cascading streams of vibrant red silk, delicately embroidered with phoenix motifs in gold thread, shimmering subtly in the humid air. Long, golden tassels adorning the four ornate rooftop corners sway gently with the sedan's steady movement. Tall, slender, emerald green bamboos form a natural barrier, a thick, milky mist swirling between the stalks. Soft, diffused light penetrates the gaps between the mist and the bamboo leaves, casting visible beams of light in the air and illuminating the floating dust particles. The sedan's base is approximately one foot off the ground, unsupported by any visible support. It drifts forward with near-silence, steady and deliberate, leaving the thick, damp, fallen leaves on the ground untouched. The entire picture is dominated by the cool deep green of the bamboo forest and the grayish white of the mist, while the fiery red and brilliant gold of the sedan chair become a strong visual center, together creating a solemn atmosphere that is quiet, mysterious and full of oriental classical fantasy.", + "seed": 6899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_0/index_1911_seed_6899_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": -0.053773327451199296, + "vlm_logprob": -0.053773327451199296, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.053773327451199296, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.16474035382270813 + ], + "opposite_logprob_tokens": [ + -1.9147403240203857 + ], + "logprob_reward": -0.16474035382270813, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16474035382270813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20745345950126648, + 0.08608914166688919, + 0.050235264003276825, + 0.09421456605195999, + 0.04492092877626419, + 0.07858498394489288, + 0.025086944922804832, + 0.24590107798576355, + 0.03641730919480324, + 0.13109630346298218 + ], + "credit_map": [ + 0.025877976790070534, + 0.025262026116251945, + 0.05451909080147743, + 0.021898146718740463, + 0.02324925921857357, + 0.018646249547600746, + 0.018705105409026146, + 0.019295601174235344, + 0.010476255789399147, + 0.00433568237349391, + 0.03054877743124962, + 0.009300528094172478, + 0.004323258530348539, + 0.005437023006379604, + 0.0056813363917171955, + 0.015986286103725433, + 0.005857021082192659, + 0.005759746767580509, + 0.0064058261923491955, + 0.0097569040954113, + 0.006539612077176571, + 0.0037503668572753668, + 0.007343345321714878, + 0.0048224423080682755, + 0.01250156294554472, + 0.006726119667291641, + 0.025049054995179176, + 0.007692619692534208, + 0.0097116157412529, + 0.005803496576845646, + 0.008343464694917202, + 0.018386630341410637, + 0.005632768850773573, + 0.007439798675477505, + 0.00627285148948431, + 0.005181537475436926, + 0.006049671210348606, + 0.004877223167568445, + 0.005003759637475014, + 0.0044633192010223866, + 0.009546677581965923, + 0.007868367247283459, + 0.010140195488929749, + 0.006581794936209917, + 0.00875888206064701, + 0.011978224851191044, + 0.01525925938040018, + 0.008451581932604313, + 0.002400557743385434, + 0.002057488076388836, + 0.004556329920887947, + 0.0021949978545308113, + 0.005222848616540432, + 0.0031304520089179277, + 0.0026025723200291395, + 0.0029216986149549484, + 0.017359068617224693, + 0.02752721682190895, + 0.017829718068242073, + 0.020896077156066895, + 0.04716319590806961, + 0.03703412041068077, + 0.04149610921740532, + 0.036595556885004044, + 0.005820016842335463, + 0.0033210869878530502, + 0.004099833779036999, + 0.00368808233179152, + 0.003545623505488038, + 0.005555151030421257, + 0.005370719358325005, + 0.005016795825213194, + 0.008043193258345127, + 0.01309734582901001, + 0.023089192807674408, + 0.013182707130908966, + 0.010360704734921455, + 0.013558612205088139, + 0.01949097029864788, + 0.030273575335741043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.16474035382270813 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.029822619631886482 + ], + "opposite_logprob_tokens": [ + -3.529822587966919 + ], + "logprob_reward": -0.029822619631886482, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822619631886482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18905548751354218, + 0.19748735427856445, + 0.05917135253548622, + 0.1060563400387764, + 0.05212860181927681, + 0.06526482850313187, + 0.07184512913227081, + 0.07618728280067444, + 0.05618296563625336, + 0.1266206055879593 + ], + "credit_map": [ + 0.02367364801466465, + 0.019334036856889725, + 0.045192230492830276, + 0.026746584102511406, + 0.01506307814270258, + 0.01930556818842888, + 0.016677340492606163, + 0.023062989115715027, + 0.018963642418384552, + 0.014182054437696934, + 0.06099621579051018, + 0.026378391310572624, + 0.009779366664588451, + 0.018283741548657417, + 0.012368584983050823, + 0.036535363644361496, + 0.008361776359379292, + 0.006780977826565504, + 0.007471098098903894, + 0.007683005183935165, + 0.0066521381959319115, + 0.0051567768678069115, + 0.01071983017027378, + 0.006345749367028475, + 0.015934105962514877, + 0.009959843009710312, + 0.014287007972598076, + 0.01042932365089655, + 0.016906769946217537, + 0.00865537952631712, + 0.012321881018579006, + 0.017562024295330048, + 0.00721338065341115, + 0.0071045150980353355, + 0.007857570424675941, + 0.006849843543022871, + 0.005461522378027439, + 0.005781104788184166, + 0.006666607689112425, + 0.005194059107452631, + 0.009565528482198715, + 0.007361957803368568, + 0.008426272310316563, + 0.005700708832591772, + 0.0071495468728244305, + 0.011400282382965088, + 0.007637666538357735, + 0.008022863417863846, + 0.007072083186358213, + 0.007218550890684128, + 0.00939427874982357, + 0.007752957288175821, + 0.014460197649896145, + 0.00830388069152832, + 0.009600386954843998, + 0.008042797446250916, + 0.008111869916319847, + 0.011259314604103565, + 0.0073856390081346035, + 0.007394432555884123, + 0.011605136096477509, + 0.01011167373508215, + 0.011117029003798962, + 0.009202186018228531, + 0.006737083662301302, + 0.005836830008774996, + 0.00658569298684597, + 0.00529780937358737, + 0.008174657821655273, + 0.007618721574544907, + 0.008912208490073681, + 0.007019962649792433, + 0.008657632395625114, + 0.011873528361320496, + 0.025901813060045242, + 0.011686149053275585, + 0.0085706552490592, + 0.012349972501397133, + 0.015457096509635448, + 0.032123763114213943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.029822619631886482 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.03812828287482262 + ], + "opposite_logprob_tokens": [ + -3.28812837600708 + ], + "logprob_reward": -0.03812828287482262, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812828287482262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15098215639591217, + 0.1875108927488327, + 0.1403675079345703, + 0.22647491097450256, + 0.046492982655763626, + 0.031108276918530464, + 0.06294088810682297, + 0.05440055578947067, + 0.03558406978845596, + 0.06413774937391281 + ], + "credit_map": [ + 0.017552200704813004, + 0.01464222464710474, + 0.03937990218400955, + 0.018115755170583725, + 0.014697647653520107, + 0.014566048979759216, + 0.012243886478245258, + 0.01978449337184429, + 0.022476976737380028, + 0.01882448047399521, + 0.04686856269836426, + 0.02840517833828926, + 0.01211401168256998, + 0.013336355797946453, + 0.01767994649708271, + 0.027805376797914505, + 0.01462307944893837, + 0.019131487235426903, + 0.020905176177620888, + 0.01920117810368538, + 0.02114901877939701, + 0.009107150137424469, + 0.026425451040267944, + 0.009824967943131924, + 0.024881158024072647, + 0.023125059902668, + 0.036438003182411194, + 0.025673791766166687, + 0.04330596700310707, + 0.017904356122016907, + 0.0322694256901741, + 0.02287713810801506, + 0.009922674857079983, + 0.004145442973822355, + 0.007817985489964485, + 0.004335382487624884, + 0.005298107396811247, + 0.004099510610103607, + 0.003784371539950371, + 0.007089510094374418, + 0.0062758298590779305, + 0.003414435777813196, + 0.004069468472152948, + 0.0027401456609368324, + 0.004488183651119471, + 0.0034545757807791233, + 0.003747791051864624, + 0.002917846664786339, + 0.006954530254006386, + 0.006778048351407051, + 0.00856622215360403, + 0.007169062737375498, + 0.009371002204716206, + 0.005755395628511906, + 0.012338949367403984, + 0.006007675547152758, + 0.006114441901445389, + 0.0056672170758247375, + 0.005401499569416046, + 0.0054121315479278564, + 0.00904123205691576, + 0.006761721335351467, + 0.008702095597982407, + 0.0073002176359295845, + 0.003972378559410572, + 0.0028397422283887863, + 0.004012813325971365, + 0.0033918777480721474, + 0.004529906902462244, + 0.005106485914438963, + 0.0066487607546150684, + 0.005082104355096817, + 0.0054199909791350365, + 0.006440532393753529, + 0.007835431955754757, + 0.006303740665316582, + 0.007986556738615036, + 0.007516952697187662, + 0.013466635718941689, + 0.009167906828224659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.03812828287482262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.0024874242953956127 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024874242953956127, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024874242953956127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1416945606470108, + 0.3618524670600891, + 0.030381323769688606, + 0.1814243495464325, + 0.08465554565191269, + 0.03412679210305214, + 0.07305775582790375, + 0.03710193559527397, + 0.02175191417336464, + 0.03395329788327217 + ], + "credit_map": [ + 0.01686919666826725, + 0.01572444662451744, + 0.026929480955004692, + 0.025611678138375282, + 0.010832835920155048, + 0.015436719171702862, + 0.01327695231884718, + 0.01701326109468937, + 0.04744401201605797, + 0.018430352210998535, + 0.13245320320129395, + 0.04026839882135391, + 0.013140780851244926, + 0.02473011054098606, + 0.02054557204246521, + 0.06484004855155945, + 0.005885094869881868, + 0.0031426912173628807, + 0.00401898892596364, + 0.005069583188742399, + 0.0022159339860081673, + 0.002790011465549469, + 0.00300531554967165, + 0.0042537059634923935, + 0.016780374571681023, + 0.005993901286274195, + 0.06122288852930069, + 0.011268294416368008, + 0.009421608410775661, + 0.011581004597246647, + 0.008834377862513065, + 0.05632193759083748, + 0.018568014726042747, + 0.008531265892088413, + 0.013035743497312069, + 0.008968891575932503, + 0.008614623919129372, + 0.011841239407658577, + 0.0058507672511041164, + 0.009245007298886776, + 0.006136286538094282, + 0.0026519373059272766, + 0.0052825892344117165, + 0.0030108755454421043, + 0.002987584099173546, + 0.004257761407643557, + 0.005276360083371401, + 0.0045234039425849915, + 0.014579376205801964, + 0.007601741701364517, + 0.009228608570992947, + 0.004380825906991959, + 0.004624546971172094, + 0.012864143587648869, + 0.005792243406176567, + 0.013986271806061268, + 0.004785784520208836, + 0.0038247478660196066, + 0.006192830391228199, + 0.0033923580776900053, + 0.004052238073199987, + 0.004697449970990419, + 0.0052947537042200565, + 0.004861779045313597, + 0.004877501633018255, + 0.002404218539595604, + 0.002604632405564189, + 0.0013946681283414364, + 0.0013986098347231746, + 0.0036799367517232895, + 0.0021948227658867836, + 0.0031975258607417345, + 0.0031382699962705374, + 0.0027951549272984266, + 0.008494998328387737, + 0.0029641164001077414, + 0.0018506405176594853, + 0.0030307264532893896, + 0.0030970063526183367, + 0.008582386188209057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0024874242953956127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.033687956631183624 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.033687956631183624, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687956631183624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08495808392763138, + 0.2540512979030609, + 0.04552200064063072, + 0.21620091795921326, + 0.06567516177892685, + 0.07838519662618637, + 0.06567459553480148, + 0.07760036736726761, + 0.06379994004964828, + 0.04813239350914955 + ], + "credit_map": [ + 0.007090254686772823, + 0.008935308083891869, + 0.013696634210646152, + 0.01846862956881523, + 0.007686441298574209, + 0.00721390638500452, + 0.00997984129935503, + 0.011887065134942532, + 0.03273428976535797, + 0.01996476761996746, + 0.08551881462335587, + 0.031127559021115303, + 0.009647714905440807, + 0.01705065928399563, + 0.01419099047780037, + 0.04381648451089859, + 0.008704050444066525, + 0.004050632938742638, + 0.00681665912270546, + 0.00630221189931035, + 0.003264917293563485, + 0.005630732513964176, + 0.0035417675971984863, + 0.007211030460894108, + 0.016624702140688896, + 0.011212245561182499, + 0.06692242622375488, + 0.019747294485569, + 0.016316276043653488, + 0.00993210170418024, + 0.013930012471973896, + 0.06151586025953293, + 0.010153447277843952, + 0.006943746004253626, + 0.00981159508228302, + 0.0063846176490187645, + 0.007166772615164518, + 0.010265701450407505, + 0.006539648864418268, + 0.008409633301198483, + 0.010341936722397804, + 0.0067457775585353374, + 0.01028306782245636, + 0.006800636183470488, + 0.008704151958227158, + 0.01085188053548336, + 0.01362891960889101, + 0.011028830893337727, + 0.009478948079049587, + 0.006651733070611954, + 0.008522615768015385, + 0.007219814695417881, + 0.007302319165319204, + 0.007854166440665722, + 0.01092677190899849, + 0.007718227803707123, + 0.007877419702708721, + 0.011826765723526478, + 0.006408276967704296, + 0.006809047423303127, + 0.015262968838214874, + 0.010010717436671257, + 0.009965629316866398, + 0.009439541958272457, + 0.006099385675042868, + 0.006377111654728651, + 0.0056171235628426075, + 0.00657821586355567, + 0.010626726783812046, + 0.007770834024995565, + 0.012346881441771984, + 0.008383657783269882, + 0.004705994855612516, + 0.004486644174903631, + 0.008932478725910187, + 0.0049088746309280396, + 0.0032386884558945894, + 0.00402032257989049, + 0.007654939778149128, + 0.010184448212385178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.033687956631183624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal tracking shot, filmed from a slightly upward angle, follows an exquisitely crafted ancient sedan chair as it slowly levitates through the depths of a mist-shrouded bamboo forest. The sedan is meticulously carved from dark lacquered wood, its frame draped in cascading streams of vibrant red silk, delicately embroidered with phoenix motifs in gold thread, shimmering subtly in the humid air. Long, golden tassels adorning the four ornate rooftop corners sway gently with the sedan's steady movement. Tall, slender, emerald green bamboos form a natural barrier, a thick, milky mist swirling between the stalks. Soft, diffused light penetrates the gaps between the mist and the bamboo leaves, casting visible beams of light in the air and illuminating the floating dust particles. The sedan's base is approximately one foot off the ground, unsupported by any visible support. It drifts forward with near-silence, steady and deliberate, leaving the thick, damp, fallen leaves on the ground untouched. The entire picture is dominated by the cool deep green of the bamboo forest and the grayish white of the mist, while the fiery red and brilliant gold of the sedan chair become a strong visual center, together creating a solemn atmosphere that is quiet, mysterious and full of oriental classical fantasy.", + "seed": 6900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_0/index_1911_seed_6900_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": -0.040328630711883307, + "vlm_logprob": -0.040328630711883307, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.040328630711883307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.10482397675514221 + ], + "opposite_logprob_tokens": [ + -2.3548240661621094 + ], + "logprob_reward": -0.10482397675514221, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10482397675514221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03076423518359661, + 0.026346782222390175, + 0.12093425542116165, + 0.1273537129163742, + 0.04781530052423477, + 0.027661677449941635, + 0.29193857312202454, + 0.03509148582816124, + 0.2701549828052521, + 0.021938951686024666 + ], + "credit_map": [ + 0.0034921602346003056, + 0.0045037358067929745, + 0.0041290950030088425, + 0.004662647377699614, + 0.004395036026835442, + 0.002900027437135577, + 0.0032813420984894037, + 0.0034001918975263834, + 0.002784857526421547, + 0.002426296006888151, + 0.005255146883428097, + 0.003660020185634494, + 0.003433062694966793, + 0.001690432894974947, + 0.003514255164191127, + 0.0035827094689011574, + 0.010283096693456173, + 0.004485634155571461, + 0.04084344953298569, + 0.008528048172593117, + 0.004293165635317564, + 0.006307199131697416, + 0.0052009569481015205, + 0.04099271073937416, + 0.009366887621581554, + 0.00442799786105752, + 0.03797479346394539, + 0.011483544483780861, + 0.00610754219815135, + 0.00845207367092371, + 0.008656129240989685, + 0.04088474065065384, + 0.004051429219543934, + 0.003444534493610263, + 0.010746978223323822, + 0.004584390204399824, + 0.0054051754996180534, + 0.0046358234249055386, + 0.004610137548297644, + 0.010336833074688911, + 0.0022180965170264244, + 0.003234918462112546, + 0.003470403142273426, + 0.00450955331325531, + 0.0025946053210645914, + 0.004035858437418938, + 0.004276262130588293, + 0.0033219819888472557, + 0.028286583721637726, + 0.033796921372413635, + 0.04010936990380287, + 0.04067723825573921, + 0.02493949979543686, + 0.033888205885887146, + 0.048440489917993546, + 0.04180026799440384, + 0.0019720494747161865, + 0.0018159003229811788, + 0.0023385281674563885, + 0.0016830723034217954, + 0.0011713843559846282, + 0.01163026038557291, + 0.0019614759366959333, + 0.012518815696239471, + 0.025989538058638573, + 0.03341633826494217, + 0.02742847427725792, + 0.02748810686171055, + 0.014378819614648819, + 0.03150239214301109, + 0.043293729424476624, + 0.06665757298469543, + 0.00264945812523365, + 0.0020292820408940315, + 0.0022099786438047886, + 0.00226994464173913, + 0.0017193086678162217, + 0.0025503768119961023, + 0.0036506650503724813, + 0.004859936889261007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.10482397675514221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.043083153665065765 + ], + "opposite_logprob_tokens": [ + -3.1680831909179688 + ], + "logprob_reward": -0.043083153665065765, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043083153665065765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15669645369052887, + 0.11338936537504196, + 0.06949400156736374, + 0.07486169040203094, + 0.10242428630590439, + 0.11243046075105667, + 0.0820278599858284, + 0.08224968612194061, + 0.10785315930843353, + 0.09857307374477386 + ], + "credit_map": [ + 0.01815549097955227, + 0.01873108558356762, + 0.032538916915655136, + 0.03040904551744461, + 0.01652010716497898, + 0.009475555270910263, + 0.01610581763088703, + 0.014760424382984638, + 0.010169058106839657, + 0.011236206628382206, + 0.014048938639461994, + 0.013759474270045757, + 0.018992485478520393, + 0.011870384216308594, + 0.02014179527759552, + 0.01317100040614605, + 0.0057717314921319485, + 0.006380877923220396, + 0.01266947016119957, + 0.011239727027714252, + 0.008177467621862888, + 0.00440548500046134, + 0.010325552895665169, + 0.01052368525415659, + 0.005797252058982849, + 0.005075360648334026, + 0.016884854063391685, + 0.008476082235574722, + 0.006867107469588518, + 0.007098405621945858, + 0.008723685517907143, + 0.015938933938741684, + 0.011952069588005543, + 0.009994162246584892, + 0.01800333708524704, + 0.010183265432715416, + 0.00906219519674778, + 0.0156635120511055, + 0.009484222158789635, + 0.018081512302160263, + 0.015604362823069096, + 0.00824732519686222, + 0.02013724483549595, + 0.009063422679901123, + 0.009480799548327923, + 0.019838081672787666, + 0.010393256321549416, + 0.019665950909256935, + 0.009942314587533474, + 0.00943545252084732, + 0.011267505586147308, + 0.00959116593003273, + 0.007535693235695362, + 0.010719376616179943, + 0.012808470986783504, + 0.010727866552770138, + 0.00615444453433156, + 0.006348686292767525, + 0.009051297791302204, + 0.006707635708153248, + 0.005624330602586269, + 0.019414765760302544, + 0.008441814221441746, + 0.02050670050084591, + 0.010936884209513664, + 0.016831953078508377, + 0.013100088573992252, + 0.011115659959614277, + 0.01038476824760437, + 0.011310091242194176, + 0.01688486710190773, + 0.017288831993937492, + 0.007745144423097372, + 0.007072813808917999, + 0.010159921832382679, + 0.012415449135005474, + 0.004446166101843119, + 0.011290674097836018, + 0.02080848254263401, + 0.024634409695863724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.043083153665065765 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.023306243121623993 + ], + "opposite_logprob_tokens": [ + -3.773306131362915 + ], + "logprob_reward": -0.023306243121623993, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306243121623993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055184539407491684, + 0.09295506775379181, + 0.05722205340862274, + 0.08390287309885025, + 0.04767578840255737, + 0.07835568487644196, + 0.22899267077445984, + 0.05131741613149643, + 0.2336912304162979, + 0.07070272415876389 + ], + "credit_map": [ + 0.004994483198970556, + 0.00633647944778204, + 0.00936469528824091, + 0.01102269534021616, + 0.0062345098704099655, + 0.004532822407782078, + 0.005478460807353258, + 0.007220394443720579, + 0.006883777212351561, + 0.005993713159114122, + 0.023539908230304718, + 0.009918234311044216, + 0.009824305772781372, + 0.006351136602461338, + 0.009710588492453098, + 0.02073340117931366, + 0.003949325997382402, + 0.004384893923997879, + 0.01049801055341959, + 0.007632669992744923, + 0.0067164101637899876, + 0.006283679511398077, + 0.007162196561694145, + 0.010594869963824749, + 0.004683531355112791, + 0.004530269652605057, + 0.019732318818569183, + 0.010986180044710636, + 0.007577795069664717, + 0.006052573677152395, + 0.009482316672801971, + 0.02085788920521736, + 0.004199107177555561, + 0.0048175957053899765, + 0.00905514508485794, + 0.0055975113064050674, + 0.004490014631301165, + 0.004763188771903515, + 0.005586537998169661, + 0.009166684933006763, + 0.01037878729403019, + 0.006193968001753092, + 0.013791555538773537, + 0.011478533037006855, + 0.005361862480640411, + 0.0108218127861619, + 0.0074449339881539345, + 0.012884232215583324, + 0.025470195338129997, + 0.0331319235265255, + 0.028432345017790794, + 0.026292886584997177, + 0.020457535982131958, + 0.024399856105446815, + 0.03417487442493439, + 0.03663305565714836, + 0.00663474015891552, + 0.008567494340240955, + 0.006459587719291449, + 0.004583117086440325, + 0.004044013563543558, + 0.0090918755158782, + 0.0052130925469100475, + 0.006723496131598949, + 0.021056845784187317, + 0.0309471245855093, + 0.02485417202115059, + 0.024366172030568123, + 0.012239160016179085, + 0.025710558518767357, + 0.03764547035098076, + 0.05687172710895538, + 0.015557121485471725, + 0.00833340734243393, + 0.006869688164442778, + 0.006178062874823809, + 0.004827449098229408, + 0.010552394203841686, + 0.009618805721402168, + 0.008765791542828083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.023306243121623993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.004092176910489798 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092176910489798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092176910489798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07168316841125488, + 0.1014801561832428, + 0.0595526248216629, + 0.1097671166062355, + 0.22117264568805695, + 0.09625861048698425, + 0.14279575645923615, + 0.030517160892486572, + 0.061410076916217804, + 0.10536262392997742 + ], + "credit_map": [ + 0.0074269757606089115, + 0.009039262309670448, + 0.014582495205104351, + 0.013693338260054588, + 0.007947467267513275, + 0.005245823413133621, + 0.0062658581882715225, + 0.0074819521978497505, + 0.008905983529984951, + 0.0074312216602265835, + 0.02214895561337471, + 0.009426518343389034, + 0.012042153626680374, + 0.010111293755471706, + 0.012360719963908195, + 0.019053315743803978, + 0.004788333550095558, + 0.0027093635872006416, + 0.01674734428524971, + 0.006034634541720152, + 0.00422479584813118, + 0.00350364251062274, + 0.004795027431100607, + 0.016749486327171326, + 0.008604182861745358, + 0.004688308574259281, + 0.029949020594358444, + 0.011368749663233757, + 0.0068727824836969376, + 0.006743596401065588, + 0.009267594665288925, + 0.032272882759571075, + 0.027341870591044426, + 0.011781211011111736, + 0.05522036552429199, + 0.013425346463918686, + 0.011296190321445465, + 0.03324244171380997, + 0.011643701232969761, + 0.057221539318561554, + 0.014776446856558323, + 0.006624435540288687, + 0.015052915550768375, + 0.008133571594953537, + 0.007954057306051254, + 0.023080909624695778, + 0.0076027545146644115, + 0.013033523224294186, + 0.023668324574828148, + 0.011976253241300583, + 0.016943473368883133, + 0.014158754609525204, + 0.01069008931517601, + 0.029858097434043884, + 0.016306446865200996, + 0.019194334745407104, + 0.003271430265158415, + 0.0017630032962188125, + 0.0032448237761855125, + 0.0030627755913883448, + 0.0029837791807949543, + 0.008612439967691898, + 0.003308469196781516, + 0.004270440433174372, + 0.003812899114564061, + 0.004825437907129526, + 0.005442918743938208, + 0.005574902519583702, + 0.004275829065591097, + 0.007406789343804121, + 0.01209360919892788, + 0.017977692186832428, + 0.00833634939044714, + 0.0069952779449522495, + 0.005591251887381077, + 0.00638264836743474, + 0.00964562688022852, + 0.03687073290348053, + 0.014965963549911976, + 0.016574770212173462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.004092176910489798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.026337603107094765 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337603107094765, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337603107094765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06601575016975403, + 0.07232078164815903, + 0.08821843564510345, + 0.023287028074264526, + 0.05354057624936104, + 0.10516583174467087, + 0.1958024799823761, + 0.11962559074163437, + 0.18618664145469666, + 0.08983685821294785 + ], + "credit_map": [ + 0.005104914307594299, + 0.009269910864531994, + 0.009403425268828869, + 0.010749009437859058, + 0.009685809724032879, + 0.00548287620767951, + 0.00921220425516367, + 0.0071076094172894955, + 0.00529057439416647, + 0.0047119311057031155, + 0.01487774308770895, + 0.009034758433699608, + 0.009514977224171162, + 0.006369872018694878, + 0.010472288355231285, + 0.012048647738993168, + 0.00818865280598402, + 0.0038768183439970016, + 0.026615863665938377, + 0.008413617499172688, + 0.004556387197226286, + 0.005075614899396896, + 0.004836821462959051, + 0.026654671877622604, + 0.0018061338923871517, + 0.0016447616508230567, + 0.003741979133337736, + 0.002435041358694434, + 0.003915147855877876, + 0.002271715085953474, + 0.003618282964453101, + 0.0038539671804755926, + 0.00569124286994338, + 0.0038777466397732496, + 0.0128624951466918, + 0.004856101702898741, + 0.00505830766633153, + 0.003941518720239401, + 0.0045345076359808445, + 0.01271866075694561, + 0.01701456494629383, + 0.009452602826058865, + 0.019629711285233498, + 0.013345787301659584, + 0.007079258095473051, + 0.01114571187645197, + 0.008487407118082047, + 0.019010799005627632, + 0.021819835528731346, + 0.03213168680667877, + 0.02364410273730755, + 0.02234635502099991, + 0.018208811059594154, + 0.020876361057162285, + 0.02516443468630314, + 0.03161089867353439, + 0.010515064932405949, + 0.005238007754087448, + 0.010489295236766338, + 0.005560978315770626, + 0.006698413286358118, + 0.031559478491544724, + 0.008207774721086025, + 0.041356585919857025, + 0.017839090898633003, + 0.027001824229955673, + 0.02167545072734356, + 0.020555613562464714, + 0.011436816304922104, + 0.019761759787797928, + 0.027047978714108467, + 0.040868129581213, + 0.010468395426869392, + 0.009519762359559536, + 0.010113945230841637, + 0.009908619336783886, + 0.00777451554313302, + 0.017709963023662567, + 0.011280134320259094, + 0.013061529025435448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.026337603107094765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal tracking shot, filmed from a slightly upward angle, follows an exquisitely crafted ancient sedan chair as it slowly levitates through the depths of a mist-shrouded bamboo forest. The sedan is meticulously carved from dark lacquered wood, its frame draped in cascading streams of vibrant red silk, delicately embroidered with phoenix motifs in gold thread, shimmering subtly in the humid air. Long, golden tassels adorning the four ornate rooftop corners sway gently with the sedan's steady movement. Tall, slender, emerald green bamboos form a natural barrier, a thick, milky mist swirling between the stalks. Soft, diffused light penetrates the gaps between the mist and the bamboo leaves, casting visible beams of light in the air and illuminating the floating dust particles. The sedan's base is approximately one foot off the ground, unsupported by any visible support. It drifts forward with near-silence, steady and deliberate, leaving the thick, damp, fallen leaves on the ground untouched. The entire picture is dominated by the cool deep green of the bamboo forest and the grayish white of the mist, while the fiery red and brilliant gold of the sedan chair become a strong visual center, together creating a solemn atmosphere that is quiet, mysterious and full of oriental classical fantasy.", + "seed": 6901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_0/index_1911_seed_6901_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": -0.043794219428673384, + "vlm_logprob": -0.043794219428673384, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.043794219428673384, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.1060991883277893 + ], + "opposite_logprob_tokens": [ + -2.3560991287231445 + ], + "logprob_reward": -0.1060991883277893, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1060991883277893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41993504762649536, + 0.046703040599823, + 0.046300116926431656, + 0.1300622969865799, + 0.034317318350076675, + 0.05003334954380989, + 0.12310104072093964, + 0.10436839610338211, + 0.01998656988143921, + 0.02519284375011921 + ], + "credit_map": [ + 0.008198973722755909, + 0.009166684001684189, + 0.012349499389529228, + 0.016301095485687256, + 0.0684283897280693, + 0.13078923523426056, + 0.0430573932826519, + 0.13164377212524414, + 0.00435026315972209, + 0.004702186677604914, + 0.004473833367228508, + 0.006944933906197548, + 0.006682143546640873, + 0.006135096307843924, + 0.0055905538611114025, + 0.00782402977347374, + 0.006392782088369131, + 0.005234141834080219, + 0.005791600793600082, + 0.006142024416476488, + 0.005227560177445412, + 0.0063558476977050304, + 0.005988618824630976, + 0.0051675401628017426, + 0.011953029781579971, + 0.01163513958454132, + 0.019376762211322784, + 0.018960587680339813, + 0.02025500498712063, + 0.01926519349217415, + 0.015484324656426907, + 0.01313226018100977, + 0.003846977138891816, + 0.003785319160670042, + 0.004908944945782423, + 0.004747207276523113, + 0.0028577190823853016, + 0.0031942755449563265, + 0.00622883765026927, + 0.004748036619275808, + 0.006956991273909807, + 0.006309030577540398, + 0.00768784387037158, + 0.006927873473614454, + 0.004950213711708784, + 0.004462119657546282, + 0.006906863301992416, + 0.0058324141427874565, + 0.009097527712583542, + 0.005968986079096794, + 0.017339205369353294, + 0.010150409303605556, + 0.038690634071826935, + 0.010699139907956123, + 0.008442170917987823, + 0.022712958976626396, + 0.005041792057454586, + 0.004024023190140724, + 0.011763077229261398, + 0.005797718185931444, + 0.030834585428237915, + 0.009299086406826973, + 0.0083440151065588, + 0.029264096170663834, + 0.001399543252773583, + 0.001014907262288034, + 0.0032027026172727346, + 0.0015301455277949572, + 0.002519124886021018, + 0.0028736512176692486, + 0.004145040642470121, + 0.0033014530781656504, + 0.003342421492561698, + 0.0020655610132962465, + 0.006929438561201096, + 0.002396712778136134, + 0.0023654981050640345, + 0.002021433087065816, + 0.002485476667061448, + 0.003586302511394024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.1060991883277893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.033725760877132416 + ], + "opposite_logprob_tokens": [ + -3.4087257385253906 + ], + "logprob_reward": -0.033725760877132416, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033725760877132416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21349205076694489, + 0.06374114006757736, + 0.0797196626663208, + 0.09532146900892258, + 0.020047850906848907, + 0.03374635800719261, + 0.26694849133491516, + 0.1774323433637619, + 0.023164570331573486, + 0.026386013254523277 + ], + "credit_map": [ + 0.011422647163271904, + 0.009929890744388103, + 0.013069894164800644, + 0.021702677011489868, + 0.04113205894827843, + 0.0619962178170681, + 0.013736859895288944, + 0.04050179943442345, + 0.005798208061605692, + 0.005165402777493, + 0.004954894073307514, + 0.006639152765274048, + 0.007662494201213121, + 0.011761766858398914, + 0.00836069043725729, + 0.013398532755672932, + 0.006139381788671017, + 0.0050643440335989, + 0.004821007139980793, + 0.005108249839395285, + 0.011865840293467045, + 0.009578334167599678, + 0.01868405193090439, + 0.018458452075719833, + 0.006854683626443148, + 0.009125478565692902, + 0.0169533621519804, + 0.01169264130294323, + 0.013476893305778503, + 0.01256285049021244, + 0.012954076752066612, + 0.011701485142111778, + 0.0015850324416533113, + 0.0015995921567082405, + 0.0023076802026480436, + 0.0020011283922940493, + 0.0019967053085565567, + 0.0022743616718798876, + 0.004633476957678795, + 0.003649874357506633, + 0.0024205667432397604, + 0.002286538016051054, + 0.0032226145267486572, + 0.002589981071650982, + 0.007858378812670708, + 0.004412399604916573, + 0.005175347905606031, + 0.005780532490462065, + 0.019762979820370674, + 0.012482869438827038, + 0.030100101605057716, + 0.01853267103433609, + 0.08986203372478485, + 0.025701580569148064, + 0.01743726059794426, + 0.05306899547576904, + 0.010237723588943481, + 0.007907354272902012, + 0.02362675964832306, + 0.012196837924420834, + 0.0487564355134964, + 0.013785775750875473, + 0.013334685936570168, + 0.047586776316165924, + 0.0020762772765010595, + 0.001722073182463646, + 0.002659060060977936, + 0.002239797031506896, + 0.003757983213290572, + 0.0034526607487350702, + 0.0034035618882626295, + 0.003853158326819539, + 0.002953361487016082, + 0.0021177881862968206, + 0.005421600770205259, + 0.002992674708366394, + 0.003464202396571636, + 0.0029890553560107946, + 0.0029519409872591496, + 0.0034953916911035776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.033725760877132416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.020594803616404533 + ], + "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": [ + 0.30044427514076233, + 0.09248588234186172, + 0.08100248128175735, + 0.06690386682748795, + 0.022580374032258987, + 0.034486863762140274, + 0.1742786318063736, + 0.17775265872478485, + 0.023988783359527588, + 0.02607620134949684 + ], + "credit_map": [ + 0.006320931948721409, + 0.0056318738497793674, + 0.005179085768759251, + 0.009327852167189121, + 0.04471858590841293, + 0.09479012340307236, + 0.03102567233145237, + 0.10345012694597244, + 0.004155687056481838, + 0.004498200956732035, + 0.0032286588102579117, + 0.005566911771893501, + 0.029529688879847527, + 0.017079319804906845, + 0.00754583440721035, + 0.02088157646358013, + 0.006057401653379202, + 0.006812822539359331, + 0.005119966343045235, + 0.0059233298525214195, + 0.019848620519042015, + 0.008397682569921017, + 0.02190222218632698, + 0.006940431892871857, + 0.007087624166160822, + 0.006436419673264027, + 0.010231572203338146, + 0.01062259916216135, + 0.007621996570378542, + 0.009590684436261654, + 0.007601760793477297, + 0.007711211685091257, + 0.002075792523100972, + 0.001858444302342832, + 0.002601268468424678, + 0.0029267147183418274, + 0.0024179338943213224, + 0.0029426803812384605, + 0.004397682845592499, + 0.003359857015311718, + 0.0035498980432748795, + 0.002580509288236499, + 0.0035841611679643393, + 0.003329402767121792, + 0.006240507122129202, + 0.005613731220364571, + 0.004935033153742552, + 0.004653621930629015, + 0.010932151228189468, + 0.006557074841111898, + 0.021879196166992188, + 0.011866440065205097, + 0.05966882035136223, + 0.016897838562726974, + 0.011248476803302765, + 0.03522862493991852, + 0.009386342950165272, + 0.007014453876763582, + 0.025802304968237877, + 0.011217613704502583, + 0.047936271876096725, + 0.01546348724514246, + 0.013811178505420685, + 0.04712101072072983, + 0.00286250957287848, + 0.0020056020002812147, + 0.005325789097696543, + 0.002974084345623851, + 0.0018616518937051296, + 0.0020283712074160576, + 0.003222241997718811, + 0.0037085332442075014, + 0.0031730036716908216, + 0.0025339561980217695, + 0.005824341904371977, + 0.0028573721647262573, + 0.0023235890548676252, + 0.0022738100960850716, + 0.0036338053178042173, + 0.003456322243437171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.020594803616404533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.0036140859592705965 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036140859592705965, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036140859592705965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21633142232894897, + 0.16212700307369232, + 0.04459979012608528, + 0.05753660574555397, + 0.027620188891887665, + 0.04225544631481171, + 0.17306649684906006, + 0.16074533760547638, + 0.06828335672616959, + 0.04743437469005585 + ], + "credit_map": [ + 0.017401952296495438, + 0.01623987779021263, + 0.014679004438221455, + 0.019060414284467697, + 0.05836865305900574, + 0.044627897441387177, + 0.02165699005126953, + 0.024296646937727928, + 0.010103479959070683, + 0.010325797833502293, + 0.010468200780451298, + 0.014965704642236233, + 0.026956994086503983, + 0.02295805886387825, + 0.02073071338236332, + 0.04561804607510567, + 0.003327433718368411, + 0.003197247628122568, + 0.002264130162075162, + 0.0028723564464598894, + 0.010734367184340954, + 0.007353619206696749, + 0.007357042748481035, + 0.00749359093606472, + 0.0029346568044275045, + 0.0027017679531127214, + 0.0031216286588460207, + 0.00530586251989007, + 0.01079641468822956, + 0.013122799806296825, + 0.007086863741278648, + 0.012466608546674252, + 0.0025128733832389116, + 0.0025680584367364645, + 0.002611799631267786, + 0.0025430095847696066, + 0.0046311030164361, + 0.0044778273440897465, + 0.004397586453706026, + 0.0038779291789978743, + 0.002718672389164567, + 0.0029286197386682034, + 0.0027810230385512114, + 0.003462420543655753, + 0.008180203847587109, + 0.009000820107758045, + 0.006243478506803513, + 0.006940209772437811, + 0.013658327981829643, + 0.007233012933284044, + 0.023246223106980324, + 0.011047852225601673, + 0.05677688494324684, + 0.01647438481450081, + 0.010996159166097641, + 0.033633653074502945, + 0.009303792379796505, + 0.006771834101527929, + 0.022071052342653275, + 0.010161872953176498, + 0.04415453225374222, + 0.012880194932222366, + 0.012275281362235546, + 0.043126773089170456, + 0.0031748113688081503, + 0.0039987703785300255, + 0.004305748268961906, + 0.006408128887414932, + 0.011747843585908413, + 0.012675531208515167, + 0.01444907858967781, + 0.011523444205522537, + 0.0031001511961221695, + 0.0031096618622541428, + 0.004189622588455677, + 0.005153708625584841, + 0.006880627013742924, + 0.0093325674533844, + 0.008580208756029606, + 0.007087827660143375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0036140859592705965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.05493725836277008 + ], + "opposite_logprob_tokens": [ + -2.9299373626708984 + ], + "logprob_reward": -0.05493725836277008, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493725836277008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5541538596153259, + 0.031441282480955124, + 0.10274722427129745, + 0.062469590455293655, + 0.021208375692367554, + 0.04328177124261856, + 0.08460231125354767, + 0.07591591030359268, + 0.010093261487782001, + 0.014086449518799782 + ], + "credit_map": [ + 0.01589268259704113, + 0.015296599827706814, + 0.019252441823482513, + 0.030760979279875755, + 0.09238914400339127, + 0.1683126837015152, + 0.051647599786520004, + 0.16060172021389008, + 0.002000633627176285, + 0.002032542834058404, + 0.0029773456044495106, + 0.004417051561176777, + 0.008198246359825134, + 0.005120542831718922, + 0.002825595438480377, + 0.0038693223614245653, + 0.004631733056157827, + 0.004013463854789734, + 0.004590322729200125, + 0.006224364973604679, + 0.026024989783763885, + 0.007617922965437174, + 0.03568781912326813, + 0.01395661011338234, + 0.004116009455174208, + 0.0054540932178497314, + 0.004973198752850294, + 0.007236017379909754, + 0.01067992951720953, + 0.009261918254196644, + 0.010218774899840355, + 0.0105296541005373, + 0.0011223581386730075, + 0.0009829289047047496, + 0.0009835916571319103, + 0.0013617112999781966, + 0.0024966145865619183, + 0.004182395059615374, + 0.006211481522768736, + 0.0038672953378409147, + 0.0025438894517719746, + 0.0028583211824297905, + 0.0022227931767702103, + 0.0031912000849843025, + 0.012219591066241264, + 0.006762760225683451, + 0.005306159146130085, + 0.008177056908607483, + 0.005511138122528791, + 0.0034879567101597786, + 0.009630807675421238, + 0.005957898683845997, + 0.0290724728256464, + 0.008390326052904129, + 0.0058480254374444485, + 0.016703683882951736, + 0.004159265663474798, + 0.0033121039159595966, + 0.009139636531472206, + 0.004677035380154848, + 0.022107629105448723, + 0.005725919269025326, + 0.005932589992880821, + 0.020861729979515076, + 0.0013469216646626592, + 0.0008004555129446089, + 0.0017692465335130692, + 0.00113850983325392, + 0.0009004988241940737, + 0.001129385782405734, + 0.0017008633585646749, + 0.0013073794543743134, + 0.0015088430372998118, + 0.0015844133449718356, + 0.0017017540521919727, + 0.0022145365364849567, + 0.0020965966396033764, + 0.0015250199940055609, + 0.001819928758777678, + 0.0016353573882952332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.05493725836277008 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, horizontal tracking shot, filmed from a slightly upward angle, follows an exquisitely crafted ancient sedan chair as it slowly levitates through the depths of a mist-shrouded bamboo forest. The sedan is meticulously carved from dark lacquered wood, its frame draped in cascading streams of vibrant red silk, delicately embroidered with phoenix motifs in gold thread, shimmering subtly in the humid air. Long, golden tassels adorning the four ornate rooftop corners sway gently with the sedan's steady movement. Tall, slender, emerald green bamboos form a natural barrier, a thick, milky mist swirling between the stalks. Soft, diffused light penetrates the gaps between the mist and the bamboo leaves, casting visible beams of light in the air and illuminating the floating dust particles. The sedan's base is approximately one foot off the ground, unsupported by any visible support. It drifts forward with near-silence, steady and deliberate, leaving the thick, damp, fallen leaves on the ground untouched. The entire picture is dominated by the cool deep green of the bamboo forest and the grayish white of the mist, while the fiery red and brilliant gold of the sedan chair become a strong visual center, together creating a solemn atmosphere that is quiet, mysterious and full of oriental classical fantasy.", + "seed": 6902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_0/index_1911_seed_6902_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": -0.07794540314935147, + "vlm_logprob": -0.07794540314935147, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.07794540314935147, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.2590671479701996 + ], + "opposite_logprob_tokens": [ + -1.509067177772522 + ], + "logprob_reward": -0.2590671479701996, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2590671479701996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2767055332660675, + 0.05781254544854164, + 0.08244480937719345, + 0.09220720082521439, + 0.05527487397193909, + 0.07916531711816788, + 0.09974358975887299, + 0.1254512518644333, + 0.0560288168489933, + 0.07516603916883469 + ], + "credit_map": [ + 0.02903018519282341, + 0.03656052052974701, + 0.0410926528275013, + 0.04676033556461334, + 0.03098115883767605, + 0.026183627545833588, + 0.037968818098306656, + 0.02812824957072735, + 0.004799208603799343, + 0.006406721659004688, + 0.005828645080327988, + 0.00876571424305439, + 0.007307905238121748, + 0.003406199859455228, + 0.013461768627166748, + 0.007836381904780865, + 0.01031417679041624, + 0.013597219251096249, + 0.007891090586781502, + 0.014674801379442215, + 0.011168847791850567, + 0.006504535209387541, + 0.00918242521584034, + 0.009111713618040085, + 0.00923941656947136, + 0.013714629225432873, + 0.009029929526150227, + 0.015429128892719746, + 0.01086763758212328, + 0.014527558349072933, + 0.009476859122514725, + 0.009922039695084095, + 0.0058648958802223206, + 0.006451147608458996, + 0.008079770021140575, + 0.009073733352124691, + 0.005313976667821407, + 0.006589548196643591, + 0.0075092739425599575, + 0.006392527837306261, + 0.012254721485078335, + 0.00874774344265461, + 0.00766737712547183, + 0.01488924864679575, + 0.007304003927856684, + 0.009698526933789253, + 0.007661012466996908, + 0.010942677967250347, + 0.014827447943389416, + 0.012925190851092339, + 0.020674938336014748, + 0.012248140759766102, + 0.006115071941167116, + 0.012527872808277607, + 0.007563668303191662, + 0.012861261144280434, + 0.00789856817573309, + 0.00865344051271677, + 0.01612422615289688, + 0.009687632322311401, + 0.02254696562886238, + 0.025859372690320015, + 0.01578948274254799, + 0.01889156363904476, + 0.00342509220354259, + 0.004511113744229078, + 0.008990833535790443, + 0.004773805849254131, + 0.006200536619871855, + 0.013204392045736313, + 0.006245813798159361, + 0.00867722649127245, + 0.008482803590595722, + 0.008114307187497616, + 0.007181998807936907, + 0.01126084290444851, + 0.008632458746433258, + 0.011437219567596912, + 0.01040908508002758, + 0.009647323749959469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.2590671479701996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.0337323322892189 + ], + "opposite_logprob_tokens": [ + -3.4087324142456055 + ], + "logprob_reward": -0.0337323322892189, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337323322892189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19418814778327942, + 0.3089542090892792, + 0.092100128531456, + 0.044680651277303696, + 0.028657162562012672, + 0.036593154072761536, + 0.070278599858284, + 0.11369827389717102, + 0.05101620405912399, + 0.05983356013894081 + ], + "credit_map": [ + 0.020008370280265808, + 0.026783209294080734, + 0.025299765169620514, + 0.03707900643348694, + 0.02160428836941719, + 0.015011445619165897, + 0.02567807026207447, + 0.022723978385329247, + 0.029269544407725334, + 0.03969919681549072, + 0.036088977009058, + 0.04832741990685463, + 0.036340635269880295, + 0.020512184128165245, + 0.05753301456570625, + 0.04118320718407631, + 0.008165393956005573, + 0.014236330054700375, + 0.014046561904251575, + 0.019931133836507797, + 0.006827536504715681, + 0.0053559853695333, + 0.01424618624150753, + 0.009290995076298714, + 0.004424251616001129, + 0.007909432053565979, + 0.004896060563623905, + 0.008120718412101269, + 0.005388037767261267, + 0.005034800618886948, + 0.004617164842784405, + 0.004290183074772358, + 0.002866484457626939, + 0.0035359119065105915, + 0.0032672742381691933, + 0.005382828880101442, + 0.003685929113999009, + 0.003228267887607217, + 0.0034182642120867968, + 0.0032721986062824726, + 0.004476935137063265, + 0.002785471500828862, + 0.002855282975360751, + 0.0035674686077982187, + 0.004631546791642904, + 0.00847602728754282, + 0.004256306681782007, + 0.005544111132621765, + 0.009936031885445118, + 0.007533773314207792, + 0.008078902959823608, + 0.008030788972973824, + 0.009306271560490131, + 0.010849852114915848, + 0.006861852016299963, + 0.009681124240159988, + 0.007176103536039591, + 0.005053548142313957, + 0.010264663957059383, + 0.006335564889013767, + 0.024859635159373283, + 0.025429265573620796, + 0.01575773023068905, + 0.018821746110916138, + 0.004698421806097031, + 0.0057811555452644825, + 0.00543208047747612, + 0.01082014013081789, + 0.00517194252461195, + 0.006011192686855793, + 0.0067408084869384766, + 0.006360457744449377, + 0.005715894978493452, + 0.010281308554112911, + 0.005174729973077774, + 0.009934907779097557, + 0.006756687071174383, + 0.006947364658117294, + 0.007524782791733742, + 0.007497876416891813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0337323322892189 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.02331206575036049 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331206575036049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331206575036049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1515849381685257, + 0.07620491832494736, + 0.0884995311498642, + 0.04910891875624657, + 0.050997599959373474, + 0.06596970558166504, + 0.17080684006214142, + 0.08961698412895203, + 0.04384029284119606, + 0.21337029337882996 + ], + "credit_map": [ + 0.013834787532687187, + 0.021141231060028076, + 0.020341098308563232, + 0.02621912583708763, + 0.017864421010017395, + 0.012330430559813976, + 0.018710801377892494, + 0.02114303596317768, + 0.007992265745997429, + 0.010712447576224804, + 0.009707056917250156, + 0.014348581433296204, + 0.006931739393621683, + 0.005207245238125324, + 0.010018209926784039, + 0.011287366040050983, + 0.009923405013978481, + 0.014158183708786964, + 0.011534193530678749, + 0.015384425409138203, + 0.008805344812572002, + 0.005366991274058819, + 0.01471230573952198, + 0.008614673279225826, + 0.004467424936592579, + 0.007461303845047951, + 0.003710599150508642, + 0.005526019725948572, + 0.00677511328831315, + 0.0090511254966259, + 0.005914774723351002, + 0.006202549673616886, + 0.0046312203630805016, + 0.005423297174274921, + 0.0059144203551113605, + 0.011112529784440994, + 0.006337035913020372, + 0.005871202796697617, + 0.005241380538791418, + 0.006466508377343416, + 0.004946156404912472, + 0.005811582785099745, + 0.005931790918111801, + 0.00924774631857872, + 0.009372962638735771, + 0.014500578865408897, + 0.005991483572870493, + 0.010167397558689117, + 0.025982510298490524, + 0.022881075739860535, + 0.03115473873913288, + 0.019780024886131287, + 0.010999065823853016, + 0.020794637501239777, + 0.017449039965867996, + 0.02176573872566223, + 0.015404739417135715, + 0.01071817334741354, + 0.014021565206348896, + 0.00909386295825243, + 0.008086452260613441, + 0.012549210339784622, + 0.00860811397433281, + 0.011134861968457699, + 0.003698291489854455, + 0.005335655529052019, + 0.0042811245657503605, + 0.005032598040997982, + 0.00445875758305192, + 0.005526233930140734, + 0.008080230094492435, + 0.007427395321428776, + 0.01727694645524025, + 0.02415468730032444, + 0.012937027029693127, + 0.04269713535904884, + 0.017402909696102142, + 0.01561137568205595, + 0.04106659069657326, + 0.04222359508275986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.02331206575036049 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.0036126605700701475 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.0036126605700701475, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036126605700701475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16110827028751373, + 0.10283448547124863, + 0.14793314039707184, + 0.036203935742378235, + 0.06870751827955246, + 0.02924320288002491, + 0.12352580577135086, + 0.1436435878276825, + 0.10865166038274765, + 0.07814843207597733 + ], + "credit_map": [ + 0.019771434366703033, + 0.018583785742521286, + 0.019088122993707657, + 0.02367360144853592, + 0.01876734383404255, + 0.015244959853589535, + 0.018519099801778793, + 0.02745990641415119, + 0.011746667325496674, + 0.009357866831123829, + 0.00835634395480156, + 0.015151332132518291, + 0.01509476825594902, + 0.012380632571876049, + 0.013865387998521328, + 0.016881488263607025, + 0.014033138751983643, + 0.014298881404101849, + 0.01745750941336155, + 0.019317014142870903, + 0.02561591938138008, + 0.015409016981720924, + 0.016236240044236183, + 0.025565410032868385, + 0.004116563592106104, + 0.004945768974721432, + 0.0030037180986255407, + 0.003951238002628088, + 0.005956537090241909, + 0.005237075965851545, + 0.0036735350731760263, + 0.005319498013705015, + 0.005570721812546253, + 0.006906714290380478, + 0.00852168072015047, + 0.015713125467300415, + 0.008265060372650623, + 0.007693244144320488, + 0.007993755862116814, + 0.008043214678764343, + 0.004553021397441626, + 0.003569747321307659, + 0.003536793403327465, + 0.003845035098493099, + 0.0023925574496388435, + 0.004558824468404055, + 0.0029065837152302265, + 0.0038806407246738672, + 0.013395840302109718, + 0.011292727664113045, + 0.014765948057174683, + 0.009968589060008526, + 0.01850191131234169, + 0.02494521252810955, + 0.014706446789205074, + 0.015949133783578873, + 0.010536789894104004, + 0.007712934166193008, + 0.01505809836089611, + 0.010153654031455517, + 0.024012068286538124, + 0.03288660943508148, + 0.021129632368683815, + 0.022153813391923904, + 0.010401182807981968, + 0.01694358140230179, + 0.01651870831847191, + 0.013570399023592472, + 0.008944151923060417, + 0.012494095601141453, + 0.01196680311113596, + 0.017812736332416534, + 0.008050416596233845, + 0.006711609195917845, + 0.008535444736480713, + 0.01054496131837368, + 0.005684759002178907, + 0.010184217244386673, + 0.014459510333836079, + 0.013977517373859882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0036126605700701475 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ancient sedan chair made of dark lacquered wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair levitating approximately one foot off the ground without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sedan chair draped in vibrant red silk embroidered with gold phoenix motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the environment consist of tall emerald green bamboo stalks with thick milky mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow horizontal tracking shot filmed from a slightly upward 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": [ + -0.07000280916690826 + ], + "opposite_logprob_tokens": [ + -2.695002794265747 + ], + "logprob_reward": -0.07000280916690826, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000280916690826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10430523008108139, + 0.11844197660684586, + 0.03493115305900574, + 0.04219421371817589, + 0.07137002050876617, + 0.10744372755289078, + 0.093421071767807, + 0.23734809458255768, + 0.13967782258987427, + 0.05086662620306015 + ], + "credit_map": [ + 0.012884584255516529, + 0.01315554603934288, + 0.011509820818901062, + 0.013324379920959473, + 0.015779010951519012, + 0.011575143784284592, + 0.011832055635750294, + 0.014244685880839825, + 0.011206910014152527, + 0.018864929676055908, + 0.015230018645524979, + 0.02322261407971382, + 0.01200728490948677, + 0.008086035028100014, + 0.019632648676633835, + 0.01019152533262968, + 0.0029697532299906015, + 0.005515233613550663, + 0.005375192500650883, + 0.007832285016775131, + 0.002691282657906413, + 0.00262653362005949, + 0.004980872850865126, + 0.002939998172223568, + 0.004287802614271641, + 0.00418064882978797, + 0.0027641833294183016, + 0.0035220682621002197, + 0.006145801395177841, + 0.010912816971540451, + 0.003977762069553137, + 0.006403128616511822, + 0.005686299875378609, + 0.007032246794551611, + 0.006271633785218, + 0.009885996580123901, + 0.013563496991991997, + 0.01117168553173542, + 0.008748541586101055, + 0.009010118432343006, + 0.00837584026157856, + 0.009298920631408691, + 0.007678207475692034, + 0.015804003924131393, + 0.015481690876185894, + 0.02408132329583168, + 0.011419814079999924, + 0.015303931199014187, + 0.015437308698892593, + 0.010812885127961636, + 0.012076123617589474, + 0.011591880582273006, + 0.00755715649574995, + 0.012162101455032825, + 0.010945910587906837, + 0.01283770240843296, + 0.01281694509088993, + 0.012806836515665054, + 0.023942461237311363, + 0.018372559919953346, + 0.04654265567660332, + 0.051239389926195145, + 0.03386477753520012, + 0.03776247799396515, + 0.013274159282445908, + 0.017427872866392136, + 0.0182796698063612, + 0.02064969576895237, + 0.0125639783218503, + 0.01942647621035576, + 0.016986150294542313, + 0.021069824695587158, + 0.003957431763410568, + 0.004698379430919886, + 0.004150084685534239, + 0.006838785018771887, + 0.003091929480433464, + 0.0061722430400550365, + 0.011138212867081165, + 0.01081955898553133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.07000280916690826 + } + ], + "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/0000110/group_1/index_tensor([3424, 3424])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_tensor([3424, 3424])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f9147ff417048647be2ee254b0c842b3f6881d1b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_tensor([3424, 3424])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, full-body shot, angled slightly upward, captures a futuristic platform constructed of smooth white marble and polished metal, majestically suspended above a vast sea of clouds, against a backdrop of a deep twilight sky. A translucent train, composed entirely of countless tiny, shimmering geometric prisms, rests at the platform's edge. Its crystalline body emanates a soft, luminous white light from within, reflecting a rainbow of colors from every facet. Shot in hyperrealistic slow motion, the train slowly sets off with near-silence and silky smoothness, drifting leisurely to the right, clearing the platform's edge without any visible contact between wheels and track. As it moves, every pulse of light and diffraction of color is clearly amplified, its ever-changing brilliance reflected on the polished platform floor. The scene, rendered in exquisite detail, creates a visual spectacle of grandeur, tranquility, and profound dreaminess.", + "seed": 1008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_3424_seed_1008_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": -0.027614236623048783, + "vlm_logprob": -0.027614236623048783, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.027614236623048783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.03227132931351662 + ], + "opposite_logprob_tokens": [ + -3.657271385192871 + ], + "logprob_reward": -0.03227132931351662, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03227132931351662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11887827515602112, + 0.11726688593626022, + 0.20373943448066711, + 0.08486369252204895, + 0.05266524478793144, + 0.02707405388355255, + 0.1857416033744812, + 0.048808902502059937, + 0.12427645921707153, + 0.03668544813990593 + ], + "credit_map": [ + 0.012608412653207779, + 0.015921005979180336, + 0.007491085212677717, + 0.008594735525548458, + 0.016617128625512123, + 0.01810006983578205, + 0.014182526618242264, + 0.025363320484757423, + 0.013305672444403172, + 0.02492680214345455, + 0.04207339510321617, + 0.007120370399206877, + 0.007472160737961531, + 0.011394021101295948, + 0.005379159934818745, + 0.005595289170742035, + 0.020409971475601196, + 0.023335935547947884, + 0.0477307066321373, + 0.010674618184566498, + 0.06663966178894043, + 0.014473126269876957, + 0.011380180716514587, + 0.009095229208469391, + 0.0099982600659132, + 0.011097311042249203, + 0.007476489525288343, + 0.007354463450610638, + 0.00880399625748396, + 0.014847028069198132, + 0.01086192112416029, + 0.014424224384129047, + 0.0065874215215444565, + 0.006531506776809692, + 0.004644542001187801, + 0.003867366584017873, + 0.006551274564117193, + 0.00886617787182331, + 0.006672539282590151, + 0.008944415487349033, + 0.0029930928722023964, + 0.004546621814370155, + 0.002653316594660282, + 0.0021718398202210665, + 0.003459468949586153, + 0.003964861389249563, + 0.002865328686311841, + 0.004419523756951094, + 0.011533441953361034, + 0.07181287556886673, + 0.004743379540741444, + 0.00471526151522994, + 0.0698508769273758, + 0.008590228855609894, + 0.006360877305269241, + 0.008134661242365837, + 0.005022106226533651, + 0.005919800605624914, + 0.007558475714176893, + 0.00866395328193903, + 0.005666665732860565, + 0.006103466730564833, + 0.0056236544623970985, + 0.004250783007591963, + 0.012347161769866943, + 0.018963012844324112, + 0.008162977173924446, + 0.00997849926352501, + 0.017608439549803734, + 0.014425121247768402, + 0.014906715601682663, + 0.027884533628821373, + 0.007352248765528202, + 0.005109813064336777, + 0.003703112481161952, + 0.0023536663502454758, + 0.0051508317701518536, + 0.005986631847918034, + 0.003469669260084629, + 0.0035594720393419266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.03227132931351662 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.016076931729912758 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.016076931729912758, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076931729912758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05171981081366539, + 0.5150429606437683, + 0.06080632656812668, + 0.035024359822273254, + 0.07755431532859802, + 0.010092710144817829, + 0.10581107437610626, + 0.07759945094585419, + 0.04505571350455284, + 0.021293235942721367 + ], + "credit_map": [ + 0.0061897169798612595, + 0.006625753827393055, + 0.005148444790393114, + 0.00921710766851902, + 0.005412074271589518, + 0.0070883408188819885, + 0.006027081049978733, + 0.006011289078742266, + 0.058870505541563034, + 0.1833987832069397, + 0.15587560832500458, + 0.026831600815057755, + 0.03836427256464958, + 0.018552321940660477, + 0.012655975297093391, + 0.0204938855022192, + 0.005756429862231016, + 0.006204518489539623, + 0.015697792172431946, + 0.0037227063439786434, + 0.009762203320860863, + 0.0037863575853407383, + 0.01050024013966322, + 0.005376080051064491, + 0.003003885271027684, + 0.005354458466172218, + 0.0016963798552751541, + 0.0033581864554435015, + 0.004045347683131695, + 0.004352256190031767, + 0.0038095058407634497, + 0.009404342621564865, + 0.008119608275592327, + 0.007795056793838739, + 0.004492026753723621, + 0.0038734599947929382, + 0.00844181701540947, + 0.014113466255366802, + 0.007037672214210033, + 0.02368120476603508, + 0.0012958524748682976, + 0.0013490986311808228, + 0.0013099542120471597, + 0.001225532265380025, + 0.0012401813874021173, + 0.0011613161768764257, + 0.0013495865277945995, + 0.0011611894005909562, + 0.008753708563745022, + 0.029773389920592308, + 0.005382212810218334, + 0.004659276455640793, + 0.03403087332844734, + 0.010800635442137718, + 0.005251153372228146, + 0.007159830536693335, + 0.00899835117161274, + 0.015222892165184021, + 0.01054843608289957, + 0.009554422460496426, + 0.009472079575061798, + 0.009082047268748283, + 0.007127621676772833, + 0.007593599613755941, + 0.004899125080555677, + 0.007695258595049381, + 0.005880123935639858, + 0.0037176748737692833, + 0.005989775061607361, + 0.005305145867168903, + 0.004429384134709835, + 0.007139226421713829, + 0.004983210004866123, + 0.0026656189002096653, + 0.0024457895196974277, + 0.0017548395553603768, + 0.0031317593529820442, + 0.0028824745677411556, + 0.0017829496646299958, + 0.0016465943772345781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.016076931729912758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.020585693418979645 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585693418979645, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585693418979645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0752633810043335, + 0.29707473516464233, + 0.14733776450157166, + 0.049733687192201614, + 0.14813290536403656, + 0.04329174384474754, + 0.09494835138320923, + 0.07700923830270767, + 0.03465476259589195, + 0.03255341574549675 + ], + "credit_map": [ + 0.010600017383694649, + 0.00885926652699709, + 0.006334318779408932, + 0.013232487253844738, + 0.007662203628569841, + 0.01358130294829607, + 0.007279721554368734, + 0.0077140661887824535, + 0.032534271478652954, + 0.10099848359823227, + 0.08910660445690155, + 0.015427852980792522, + 0.021757395938038826, + 0.013500215485692024, + 0.007341568358242512, + 0.01640837825834751, + 0.014955476857721806, + 0.015663964673876762, + 0.029714170843362808, + 0.006766567472368479, + 0.04657484218478203, + 0.012507569044828415, + 0.0113894147798419, + 0.009765753522515297, + 0.0034388615749776363, + 0.007832592353224754, + 0.004067691974341869, + 0.005035229027271271, + 0.0058231046423316, + 0.00941772572696209, + 0.004444183316081762, + 0.009674293920397758, + 0.01418831292539835, + 0.013823841698467731, + 0.008825445547699928, + 0.006472955457866192, + 0.01611918956041336, + 0.02851846069097519, + 0.014607768505811691, + 0.04557691887021065, + 0.004937073681503534, + 0.005051141604781151, + 0.0057920776307582855, + 0.004356036894023418, + 0.005382146220654249, + 0.006191018503159285, + 0.005556390155106783, + 0.006025857757776976, + 0.008072884753346443, + 0.02275637537240982, + 0.00610480597242713, + 0.006410142406821251, + 0.02266327291727066, + 0.012850909493863583, + 0.008050899021327496, + 0.008039074949920177, + 0.007833623327314854, + 0.012564566917717457, + 0.00827848631888628, + 0.00847990345209837, + 0.010701564140617847, + 0.010472596623003483, + 0.008612438105046749, + 0.010066063143312931, + 0.00392315536737442, + 0.00444719148799777, + 0.0025009361561387777, + 0.0030115293338894844, + 0.006481139920651913, + 0.0058995638974010944, + 0.003716628532856703, + 0.004674619063735008, + 0.005760046187788248, + 0.005152834579348564, + 0.0030301345977932215, + 0.0029538788367062807, + 0.0046959733590483665, + 0.004501929506659508, + 0.0034637432545423508, + 0.0029948733281344175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.020585693418979645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.05494651198387146 + ], + "opposite_logprob_tokens": [ + -2.9299464225769043 + ], + "logprob_reward": -0.05494651198387146, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494651198387146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1664106547832489, + 0.14551569521427155, + 0.04932768642902374, + 0.06599147617816925, + 0.053820330649614334, + 0.03801572322845459, + 0.3894830644130707, + 0.035073716193437576, + 0.03017565980553627, + 0.026185953989624977 + ], + "credit_map": [ + 0.0163946021348238, + 0.020437365397810936, + 0.011265923269093037, + 0.017660101875662804, + 0.025535250082612038, + 0.019175032153725624, + 0.017403824254870415, + 0.038538575172424316, + 0.021053101867437363, + 0.049392811954021454, + 0.016768358647823334, + 0.007279286161065102, + 0.014779048040509224, + 0.015262216329574585, + 0.0061114770360291, + 0.014869402162730694, + 0.007162491790950298, + 0.006092451047152281, + 0.004138197284191847, + 0.0027569890953600407, + 0.012624725699424744, + 0.004730992019176483, + 0.007872243411839008, + 0.003949600271880627, + 0.007025209255516529, + 0.00880108680576086, + 0.005948766600340605, + 0.008828538469970226, + 0.006708400323987007, + 0.006579954642802477, + 0.007806108798831701, + 0.01429341733455658, + 0.0059811812825500965, + 0.0052888295613229275, + 0.0036542722955346107, + 0.003224822459742427, + 0.005394603591412306, + 0.009178857319056988, + 0.006113140378147364, + 0.01498462539166212, + 0.0038601208943873644, + 0.004413015674799681, + 0.003843600396066904, + 0.0030665092635899782, + 0.005858589429408312, + 0.005845889914780855, + 0.0043711294420063496, + 0.0067568751983344555, + 0.02421543560922146, + 0.15482331812381744, + 0.009093880653381348, + 0.007803487591445446, + 0.156478151679039, + 0.017207404598593712, + 0.01042471919208765, + 0.009436680935323238, + 0.004097359254956245, + 0.005867618601769209, + 0.00433048652485013, + 0.0037617094349116087, + 0.004607526119798422, + 0.0044822851195931435, + 0.0035115256905555725, + 0.004415207542479038, + 0.0025262890849262476, + 0.002815718064084649, + 0.0027221888303756714, + 0.0035134179051965475, + 0.004628343507647514, + 0.0046854205429553986, + 0.003865411039441824, + 0.005418873392045498, + 0.003786192275583744, + 0.004036965314298868, + 0.002385919913649559, + 0.0018806987209245563, + 0.0035472335293889046, + 0.004716489929705858, + 0.0027036364190280437, + 0.0031288210302591324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.05494651198387146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.014190716668963432 + ], + "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": [ + 0.14906643331050873, + 0.32091325521469116, + 0.13959625363349915, + 0.05907394737005234, + 0.09030665457248688, + 0.08011975139379501, + 0.021839585155248642, + 0.06000151485204697, + 0.02409408986568451, + 0.05498848482966423 + ], + "credit_map": [ + 0.017714187502861023, + 0.01687655597925186, + 0.012692316435277462, + 0.018179573118686676, + 0.01845526695251465, + 0.019270671531558037, + 0.014861810021102428, + 0.031016064807772636, + 0.037688013166189194, + 0.1099468469619751, + 0.07534104585647583, + 0.016262328252196312, + 0.02692132443189621, + 0.024401647970080376, + 0.009974189102649689, + 0.02037785016000271, + 0.012412096373736858, + 0.01513609942048788, + 0.034524351358413696, + 0.007648907136172056, + 0.03722837567329407, + 0.012675857171416283, + 0.012228044681251049, + 0.00774253299459815, + 0.0036198117304593325, + 0.010169780813157558, + 0.0042362334206700325, + 0.004656266421079636, + 0.007871253415942192, + 0.01458621583878994, + 0.006241170689463615, + 0.007693220395594835, + 0.010313821025192738, + 0.00867490191012621, + 0.00636830972507596, + 0.005762290675193071, + 0.008586213923990726, + 0.020521510392427444, + 0.008958407677710056, + 0.021121203899383545, + 0.008843849413096905, + 0.010963333770632744, + 0.0077709476463496685, + 0.007694334723055363, + 0.008475907146930695, + 0.00987529382109642, + 0.009610473178327084, + 0.01688562147319317, + 0.002437927993014455, + 0.003400125540792942, + 0.002482914598658681, + 0.001977628329768777, + 0.004346448462456465, + 0.0021242937073111534, + 0.002874799072742462, + 0.002195450011640787, + 0.005742979701608419, + 0.00716701103374362, + 0.008718320168554783, + 0.007760328706353903, + 0.007313959300518036, + 0.0073246038518846035, + 0.008489558473229408, + 0.007484752219170332, + 0.00280071166343987, + 0.0031454176642000675, + 0.0017574458615854383, + 0.002794245956465602, + 0.003467525588348508, + 0.004255595616996288, + 0.0031681377440690994, + 0.002705011982470751, + 0.010422319173812866, + 0.007765116170048714, + 0.006141721270978451, + 0.004231389611959457, + 0.007265220396220684, + 0.008200139738619328, + 0.005579870194196701, + 0.005382708739489317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.014190716668963432 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, full-body shot, angled slightly upward, captures a futuristic platform constructed of smooth white marble and polished metal, majestically suspended above a vast sea of clouds, against a backdrop of a deep twilight sky. A translucent train, composed entirely of countless tiny, shimmering geometric prisms, rests at the platform's edge. Its crystalline body emanates a soft, luminous white light from within, reflecting a rainbow of colors from every facet. Shot in hyperrealistic slow motion, the train slowly sets off with near-silence and silky smoothness, drifting leisurely to the right, clearing the platform's edge without any visible contact between wheels and track. As it moves, every pulse of light and diffraction of color is clearly amplified, its ever-changing brilliance reflected on the polished platform floor. The scene, rendered in exquisite detail, creates a visual spectacle of grandeur, tranquility, and profound dreaminess.", + "seed": 1009, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_3424_seed_1009_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": -0.018762947898358105, + "vlm_logprob": -0.018762947898358105, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.018762947898358105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.025267187505960464 + ], + "opposite_logprob_tokens": [ + -3.9002671241760254 + ], + "logprob_reward": -0.025267187505960464, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025267187505960464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03591689467430115, + 0.02602650411427021, + 0.11410205811262131, + 0.04325425252318382, + 0.06815377622842789, + 0.06036248430609703, + 0.07442391663789749, + 0.03412526473402977, + 0.025752395391464233, + 0.5178825259208679 + ], + "credit_map": [ + 0.004056737758219242, + 0.0064079733565449715, + 0.002547377021983266, + 0.0027776032220572233, + 0.007973385974764824, + 0.005989888682961464, + 0.003523485502228141, + 0.002640442457050085, + 0.0030222884379327297, + 0.00439520413056016, + 0.002748208586126566, + 0.0038524093106389046, + 0.0033733427990227938, + 0.0028654884081333876, + 0.00322536937892437, + 0.0025441888719797134, + 0.015560189262032509, + 0.02548222802579403, + 0.004673427902162075, + 0.006948193069547415, + 0.013631979003548622, + 0.03530494123697281, + 0.006449506618082523, + 0.006051592994481325, + 0.004241702612489462, + 0.012298261746764183, + 0.0037603650707751513, + 0.005029911641031504, + 0.006111679598689079, + 0.005239530000835657, + 0.003870930289849639, + 0.0027018669061362743, + 0.002389624249190092, + 0.0033319408539682627, + 0.002005066256970167, + 0.0021421059500426054, + 0.026521733030676842, + 0.025804057717323303, + 0.0023224998731166124, + 0.0036367415450513363, + 0.006162228528410196, + 0.009766420349478722, + 0.0032313414849340916, + 0.006760070100426674, + 0.012911500409245491, + 0.012172894552350044, + 0.004762292839586735, + 0.004595728125423193, + 0.008175086230039597, + 0.013805033639073372, + 0.010112046264111996, + 0.009167755022644997, + 0.008523973636329174, + 0.009833971038460732, + 0.007744041737169027, + 0.0070620025508105755, + 0.0030586058273911476, + 0.004333600867539644, + 0.0034614696633070707, + 0.005405203904956579, + 0.003479010658338666, + 0.004406406544148922, + 0.005068805068731308, + 0.004912157543003559, + 0.0020916599314659834, + 0.0030284409876912832, + 0.001785241300240159, + 0.002993400674313307, + 0.00714662903919816, + 0.004166898783296347, + 0.0024534729309380054, + 0.0020866505801677704, + 0.13211043179035187, + 0.03420035541057587, + 0.027900459244847298, + 0.04157031700015068, + 0.033361103385686874, + 0.18892624974250793, + 0.03005250357091427, + 0.02976100891828537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.025267187505960464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.007641016039997339 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007641016039997339, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641016039997339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17391788959503174, + 0.04299471154808998, + 0.127164825797081, + 0.020767871290445328, + 0.111885666847229, + 0.06199491769075394, + 0.03373926132917404, + 0.024126168340444565, + 0.03185224533081055, + 0.37155649065971375 + ], + "credit_map": [ + 0.006888280156999826, + 0.01085971761494875, + 0.004261970520019531, + 0.005122878588736057, + 0.0708027258515358, + 0.06727691739797592, + 0.00504815811291337, + 0.0036572287790477276, + 0.005321897100657225, + 0.00674925884231925, + 0.0038396709132939577, + 0.005702622700482607, + 0.005610570777207613, + 0.006970791611820459, + 0.005296803079545498, + 0.0035030925646424294, + 0.011361529119312763, + 0.02067548967897892, + 0.009853629395365715, + 0.015692682936787605, + 0.026771504431962967, + 0.015018673613667488, + 0.015948442742228508, + 0.011842858977615833, + 0.0022202241234481335, + 0.002442736644297838, + 0.002098868368193507, + 0.00279845017939806, + 0.002116951160132885, + 0.002920422237366438, + 0.003972189966589212, + 0.0021980248857289553, + 0.005955223459750414, + 0.006395203992724419, + 0.0061602299101650715, + 0.004979284480214119, + 0.03692923113703728, + 0.03680017590522766, + 0.006560145877301693, + 0.008106168359518051, + 0.00730159180238843, + 0.008880778215825558, + 0.007169733289629221, + 0.0071141161024570465, + 0.01066775992512703, + 0.012168128043413162, + 0.003930058795958757, + 0.004762748721987009, + 0.003583006327971816, + 0.004836970008909702, + 0.0032376854214817286, + 0.004442634526640177, + 0.004337893798947334, + 0.0048431046307086945, + 0.005061532370746136, + 0.0033964302856475115, + 0.0026340291369706392, + 0.0038499468937516212, + 0.0025518005713820457, + 0.004407661035656929, + 0.002456345595419407, + 0.0034529296681284904, + 0.002491010818630457, + 0.002282440662384033, + 0.0030385376885533333, + 0.00268790521658957, + 0.0026529268361628056, + 0.0031059905886650085, + 0.008992109447717667, + 0.005157718900591135, + 0.003630194580182433, + 0.00258685857988894, + 0.09808642417192459, + 0.02370939403772354, + 0.01963127963244915, + 0.029259713366627693, + 0.023137737065553665, + 0.13504824042320251, + 0.02163507044315338, + 0.021048622205853462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.007641016039997339 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.01607341133058071 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.01607341133058071, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607341133058071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05767003446817398, + 0.039451807737350464, + 0.29683712124824524, + 0.05374687537550926, + 0.059345122426748276, + 0.05221148207783699, + 0.03033345565199852, + 0.04415467754006386, + 0.02626650035381317, + 0.33998289704322815 + ], + "credit_map": [ + 0.00572672626003623, + 0.006122003775089979, + 0.006530428770929575, + 0.00448921974748373, + 0.012864816002547741, + 0.012971815653145313, + 0.004408787935972214, + 0.004556235857307911, + 0.003798204706981778, + 0.0065671768970787525, + 0.003727705916389823, + 0.005966711789369583, + 0.004202764481306076, + 0.00508952746167779, + 0.006291537545621395, + 0.0038081796374171972, + 0.04091833904385567, + 0.0696779415011406, + 0.010050580836832523, + 0.020700780674815178, + 0.032586101442575455, + 0.09685248136520386, + 0.012730005197227001, + 0.013320871628820896, + 0.005951886996626854, + 0.012797273695468903, + 0.005531278904527426, + 0.006048060022294521, + 0.006506855133920908, + 0.006721879355609417, + 0.006126404739916325, + 0.004063239321112633, + 0.003674027044326067, + 0.0038850943092256784, + 0.004098841454833746, + 0.002164060715585947, + 0.019503427669405937, + 0.020504113286733627, + 0.0030659325420856476, + 0.0024496244732290506, + 0.005292421206831932, + 0.008334217593073845, + 0.0035892254672944546, + 0.004823635332286358, + 0.011095883324742317, + 0.01250990480184555, + 0.0029104254208505154, + 0.003655769629403949, + 0.003788999281823635, + 0.005533408839255571, + 0.0026753603015094995, + 0.003314034780487418, + 0.004201514180749655, + 0.0038562684785574675, + 0.0039618718437850475, + 0.003001998644322157, + 0.004922355990856886, + 0.005426724907010794, + 0.004996795207262039, + 0.005889229942113161, + 0.004416132811456919, + 0.006388784386217594, + 0.005466907285153866, + 0.006647746078670025, + 0.0027136041317135096, + 0.0021702363155782223, + 0.003945705480873585, + 0.003393601393327117, + 0.0036208282690495253, + 0.0029977327212691307, + 0.0041017415933310986, + 0.0033230488188564777, + 0.08561081439256668, + 0.022243715822696686, + 0.01835210993885994, + 0.027886128053069115, + 0.02176729217171669, + 0.12516231834888458, + 0.019276447594165802, + 0.019684098660945892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.01607341133058071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.038097646087408066 + ], + "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": [ + 0.0466969795525074, + 0.027078326791524887, + 0.08026420325040817, + 0.018191516399383545, + 0.05076802521944046, + 0.026285842061042786, + 0.02530013956129551, + 0.04658102989196777, + 0.03598727285861969, + 0.6428465843200684 + ], + "credit_map": [ + 0.005546220578253269, + 0.007699375040829182, + 0.0037481943145394325, + 0.003483143635094166, + 0.009352223016321659, + 0.009416567161679268, + 0.003774968907237053, + 0.0036762875970453024, + 0.0029468617867678404, + 0.004590590950101614, + 0.0033214243594557047, + 0.003544524544849992, + 0.003483643289655447, + 0.0037641574162989855, + 0.002752485452219844, + 0.002674640854820609, + 0.008870354853570461, + 0.019447464495897293, + 0.0040458072908222675, + 0.006584848742932081, + 0.010722238570451736, + 0.022321507334709167, + 0.0048203496262431145, + 0.0034516381565481424, + 0.0019936321768909693, + 0.004076461773365736, + 0.0013418860035017133, + 0.001971726305782795, + 0.002477071713656187, + 0.003142582718282938, + 0.0019239153480157256, + 0.0012642431538552046, + 0.0033843237906694412, + 0.005514252465218306, + 0.004842544440180063, + 0.0034042836632579565, + 0.012230372056365013, + 0.011541446670889854, + 0.0036737637128680944, + 0.006177044007927179, + 0.002203762298449874, + 0.0030169966630637646, + 0.002638030331581831, + 0.003302978817373514, + 0.004668671637773514, + 0.0048894560895860195, + 0.0029319010209292173, + 0.0026340456679463387, + 0.0026759218890219927, + 0.003250852460041642, + 0.004014947451651096, + 0.0029605221934616566, + 0.0030493028461933136, + 0.003733449149876833, + 0.0030156446155160666, + 0.002599499188363552, + 0.005419227760285139, + 0.0065971375443041325, + 0.006008906289935112, + 0.008656525053083897, + 0.004159501288086176, + 0.005797848105430603, + 0.005232176743447781, + 0.004709708970040083, + 0.002426168415695429, + 0.00275838072411716, + 0.003845649305731058, + 0.004313951823860407, + 0.005777968093752861, + 0.004636313766241074, + 0.005520091392099857, + 0.006708754692226648, + 0.15868616104125977, + 0.04253678768873215, + 0.03720124438405037, + 0.05616850033402443, + 0.04209539294242859, + 0.2120175063610077, + 0.05511137843132019, + 0.039029642939567566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.038097646087408066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.006735478527843952 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735478527843952, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735478527843952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06143027916550636, + 0.022305166348814964, + 0.20842517912387848, + 0.032872483134269714, + 0.18102379143238068, + 0.0382215678691864, + 0.028348546475172043, + 0.050233762711286545, + 0.026982702314853668, + 0.3501565158367157 + ], + "credit_map": [ + 0.0033710156567394733, + 0.008347224444150925, + 0.004143334925174713, + 0.0029500240925699472, + 0.01846017688512802, + 0.0190843865275383, + 0.0027486616745591164, + 0.0023254596162587404, + 0.0022956524044275284, + 0.003944208845496178, + 0.0022923615761101246, + 0.003362623741850257, + 0.0025011575780808926, + 0.0023036457132548094, + 0.003174317069351673, + 0.0024312015157192945, + 0.026424238458275795, + 0.05307203158736229, + 0.008011594414710999, + 0.014361966401338577, + 0.024039147421717644, + 0.0656360313296318, + 0.009276829659938812, + 0.007603348232805729, + 0.0039044686127454042, + 0.009392512030899525, + 0.0033248683903366327, + 0.003032080363482237, + 0.004599121864885092, + 0.0043787467293441296, + 0.0019167889840900898, + 0.0023238989524543285, + 0.0052738492377102375, + 0.007510280702263117, + 0.006877537816762924, + 0.005392721388489008, + 0.07080145180225372, + 0.06885458528995514, + 0.0060700345784425735, + 0.010243351571261883, + 0.0028823420871049166, + 0.005155833438038826, + 0.005485157482326031, + 0.005135640501976013, + 0.006348832044750452, + 0.0062371972016990185, + 0.0037943520583212376, + 0.0031822146847844124, + 0.002257710322737694, + 0.003625904442742467, + 0.0024675270542502403, + 0.002415937138721347, + 0.006100472528487444, + 0.005769894924014807, + 0.003261044155806303, + 0.002450056839734316, + 0.0041808136738836765, + 0.006882991176098585, + 0.004763165023177862, + 0.007473737932741642, + 0.005368182435631752, + 0.006499430630356073, + 0.007824989035725594, + 0.007240454666316509, + 0.0020359123591333628, + 0.0023152795620262623, + 0.0030712545849382877, + 0.0037931001279503107, + 0.004803793504834175, + 0.003214598400518298, + 0.004501863848417997, + 0.003246902022510767, + 0.09280265867710114, + 0.022039663046598434, + 0.01803578995168209, + 0.02736729197204113, + 0.02176705375313759, + 0.13063780963420868, + 0.018653174862265587, + 0.01885310746729374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.006735478527843952 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, full-body shot, angled slightly upward, captures a futuristic platform constructed of smooth white marble and polished metal, majestically suspended above a vast sea of clouds, against a backdrop of a deep twilight sky. A translucent train, composed entirely of countless tiny, shimmering geometric prisms, rests at the platform's edge. Its crystalline body emanates a soft, luminous white light from within, reflecting a rainbow of colors from every facet. Shot in hyperrealistic slow motion, the train slowly sets off with near-silence and silky smoothness, drifting leisurely to the right, clearing the platform's edge without any visible contact between wheels and track. As it moves, every pulse of light and diffraction of color is clearly amplified, its ever-changing brilliance reflected on the polished platform floor. The scene, rendered in exquisite detail, creates a visual spectacle of grandeur, tranquility, and profound dreaminess.", + "seed": 1010, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_3424_seed_1010_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": -0.030468107387423515, + "vlm_logprob": -0.030468107387423515, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.030468107387423515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.019704081118106842 + ], + "opposite_logprob_tokens": [ + -4.1447038650512695 + ], + "logprob_reward": -0.019704081118106842, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019704081118106842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09713076800107956, + 0.19381269812583923, + 0.10416936129331589, + 0.15233585238456726, + 0.04311496764421463, + 0.053428877145051956, + 0.07115267217159271, + 0.03287435695528984, + 0.19131223857402802, + 0.060668252408504486 + ], + "credit_map": [ + 0.007220693398267031, + 0.017216939479112625, + 0.011012577451765537, + 0.01520578283816576, + 0.014782721176743507, + 0.017679421231150627, + 0.006422648672014475, + 0.007589979097247124, + 0.009714474901556969, + 0.036830127239227295, + 0.02157125063240528, + 0.03125995397567749, + 0.037599511444568634, + 0.03966753929853439, + 0.008569758385419846, + 0.008600066415965557, + 0.006682475097477436, + 0.023911524564027786, + 0.010486327111721039, + 0.009367188438773155, + 0.01719430834054947, + 0.024738233536481857, + 0.0055587394163012505, + 0.0062305619940161705, + 0.014737175777554512, + 0.01786801405251026, + 0.010853742249310017, + 0.010360711254179478, + 0.022836923599243164, + 0.02165958285331726, + 0.01843557506799698, + 0.03558412194252014, + 0.00529851857572794, + 0.007947453297674656, + 0.0065189385786652565, + 0.0024653421714901924, + 0.005931486841291189, + 0.007623713929206133, + 0.0029034744948148727, + 0.004426038358360529, + 0.006902921944856644, + 0.008933204226195812, + 0.008598863147199154, + 0.0039383359253406525, + 0.009407551027834415, + 0.005578029900789261, + 0.005850228946655989, + 0.004219742026180029, + 0.014229429885745049, + 0.011272724717855453, + 0.007157723885029554, + 0.005207629408687353, + 0.013152312487363815, + 0.01065003965049982, + 0.005077914800494909, + 0.004404895007610321, + 0.003834527451545, + 0.004547479562461376, + 0.003945233765989542, + 0.00520006800070405, + 0.003991230856627226, + 0.007329112384468317, + 0.002207235200330615, + 0.001819467288441956, + 0.03213263303041458, + 0.05245203524827957, + 0.011403027921915054, + 0.008950315415859222, + 0.0331137552857399, + 0.03639920800924301, + 0.005468629766255617, + 0.011392614804208279, + 0.007836814038455486, + 0.008495517075061798, + 0.005951379891484976, + 0.002448574174195528, + 0.017020029947161674, + 0.010686511173844337, + 0.0031696222722530365, + 0.005059797782450914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.019704081118106842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.009792622178792953 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792622178792953, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792622178792953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354849487543106, + 0.17973336577415466, + 0.039178431034088135, + 0.07196047157049179, + 0.023044075816869736, + 0.05549662560224533, + 0.03167843818664551, + 0.03372572362422943, + 0.3848036527633667, + 0.04489433765411377 + ], + "credit_map": [ + 0.011159409768879414, + 0.026281606405973434, + 0.019467024132609367, + 0.014922789297997952, + 0.01955297589302063, + 0.023450752720236778, + 0.010113842785358429, + 0.01053653284907341, + 0.009243146516382694, + 0.031072841957211494, + 0.026382319629192352, + 0.02320082299411297, + 0.027313046157360077, + 0.03370698541402817, + 0.011585444211959839, + 0.017228739336133003, + 0.0030150131788104773, + 0.005422153510153294, + 0.005001258570700884, + 0.004176480229943991, + 0.006412559654563665, + 0.007213447242975235, + 0.003771531628444791, + 0.004165983758866787, + 0.008163281716406345, + 0.007897092960774899, + 0.009203002788126469, + 0.005545218475162983, + 0.010188684798777103, + 0.009225740097463131, + 0.0075238170102238655, + 0.014213631860911846, + 0.0035800463519990444, + 0.0028153571765869856, + 0.0028002874460071325, + 0.0013765029143542051, + 0.0038352033589035273, + 0.004474731162190437, + 0.001524855848401785, + 0.002637089928612113, + 0.007683365605771542, + 0.009951828978955746, + 0.00800559762865305, + 0.003878555726259947, + 0.010074089281260967, + 0.006591887678951025, + 0.004481774289160967, + 0.004829524550586939, + 0.005074026994407177, + 0.005869797896593809, + 0.002737807808443904, + 0.0020574068184942007, + 0.004732191562652588, + 0.00657297857105732, + 0.0021901417057961226, + 0.0024440886918455362, + 0.005098957568407059, + 0.006671971175819635, + 0.0033934435341507196, + 0.004640151746571064, + 0.004040551837533712, + 0.0057435655035078526, + 0.002167732222005725, + 0.0019693481735885143, + 0.06357932835817337, + 0.09622237831354141, + 0.016342835500836372, + 0.021170370280742645, + 0.06913361698389053, + 0.08871717005968094, + 0.012410435825586319, + 0.017227502539753914, + 0.005786082707345486, + 0.004559589084237814, + 0.006594665348529816, + 0.0036048267502337694, + 0.00746820867061615, + 0.007155165076255798, + 0.003637331072241068, + 0.0060884649865329266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.009792622178792953 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.029801329597830772 + ], + "opposite_logprob_tokens": [ + -3.529801368713379 + ], + "logprob_reward": -0.029801329597830772, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801329597830772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10332414507865906, + 0.22257651388645172, + 0.05223814398050308, + 0.08430886268615723, + 0.0270011555403471, + 0.04905375838279724, + 0.11847270280122757, + 0.03614742308855057, + 0.20729593932628632, + 0.09958135336637497 + ], + "credit_map": [ + 0.007951470091938972, + 0.016931887716054916, + 0.012946580536663532, + 0.014902093447744846, + 0.015734996646642685, + 0.01354158017784357, + 0.00811686459928751, + 0.013198679313063622, + 0.01169212069362402, + 0.03449864685535431, + 0.051862385123968124, + 0.026976222172379494, + 0.035224057734012604, + 0.02896997705101967, + 0.01693868823349476, + 0.01641441509127617, + 0.004776954185217619, + 0.008039341308176517, + 0.008490338921546936, + 0.0061448970809578896, + 0.0076322234235703945, + 0.008884984999895096, + 0.003850810695439577, + 0.004418597090989351, + 0.009123054333031178, + 0.00906431209295988, + 0.008341344073414803, + 0.006959511432796717, + 0.010923757217824459, + 0.009715546853840351, + 0.009489715099334717, + 0.020691625773906708, + 0.004153926856815815, + 0.003877792041748762, + 0.0034083763603121042, + 0.002308183116838336, + 0.005308159161359072, + 0.0030079041607677937, + 0.002086973749101162, + 0.002849842654541135, + 0.005890567321330309, + 0.010407413356006145, + 0.003908786457031965, + 0.0037818357814103365, + 0.006881317123770714, + 0.00502057047560811, + 0.007656183559447527, + 0.005507085006684065, + 0.02228202298283577, + 0.013500047847628593, + 0.00914902612566948, + 0.008665405213832855, + 0.01820394955575466, + 0.02492632158100605, + 0.011092554777860641, + 0.010653379373252392, + 0.0044137584045529366, + 0.006116827949881554, + 0.00290201255120337, + 0.004056953825056553, + 0.005139382556080818, + 0.0076504177413880825, + 0.0027857236564159393, + 0.0030823505949229, + 0.03525136038661003, + 0.056862153112888336, + 0.01400022953748703, + 0.009222024120390415, + 0.0351945199072361, + 0.03829389810562134, + 0.006090902257710695, + 0.012380854226648808, + 0.015523853711783886, + 0.012756354175508022, + 0.006634407676756382, + 0.004299007821828127, + 0.02199539914727211, + 0.022810427471995354, + 0.005230433773249388, + 0.010331474244594574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.029801329597830772 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.05494910478591919 + ], + "opposite_logprob_tokens": [ + -2.9299490451812744 + ], + "logprob_reward": -0.05494910478591919, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494910478591919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0417165569961071, + 0.07623915374279022, + 0.05289953574538231, + 0.049211449921131134, + 0.07496284693479538, + 0.03846144676208496, + 0.12708912789821625, + 0.023082703351974487, + 0.45871299505233765, + 0.0576242059469223 + ], + "credit_map": [ + 0.0026656510308384895, + 0.007845986634492874, + 0.004840657114982605, + 0.00534803094342351, + 0.006618613377213478, + 0.007484324276447296, + 0.003079479793086648, + 0.0038338180165737867, + 0.004817788489162922, + 0.011506648734211922, + 0.018902763724327087, + 0.009122979827225208, + 0.01207984983921051, + 0.00875021517276764, + 0.005444557871669531, + 0.005614348687231541, + 0.0035159096587449312, + 0.00868542306125164, + 0.009552226401865482, + 0.005940740462392569, + 0.006819820497184992, + 0.00983990915119648, + 0.0032925386913120747, + 0.005252969916909933, + 0.006665970664471388, + 0.005443537142127752, + 0.006437825970351696, + 0.0037206753622740507, + 0.006153740454465151, + 0.005903179757297039, + 0.006777162663638592, + 0.008109360001981258, + 0.008672041818499565, + 0.009127054363489151, + 0.012664559297263622, + 0.0053575229831039906, + 0.010847889818251133, + 0.009203121066093445, + 0.0065163010731339455, + 0.012574354186654091, + 0.0057120174169540405, + 0.006184093654155731, + 0.004848194308578968, + 0.0036977340932935476, + 0.005310091655701399, + 0.003820911981165409, + 0.004116200376302004, + 0.004772201180458069, + 0.029030267149209976, + 0.01859680376946926, + 0.01336375530809164, + 0.008004942908883095, + 0.01971759833395481, + 0.018407652154564857, + 0.009831726551055908, + 0.01013638824224472, + 0.003046528436243534, + 0.004040219821035862, + 0.003186739282682538, + 0.002759899478405714, + 0.0030662319622933865, + 0.0038553287740796804, + 0.0014841299271211028, + 0.0016436284640803933, + 0.07819883525371552, + 0.1202630028128624, + 0.02174478955566883, + 0.022256018593907356, + 0.08145462721586227, + 0.09983900189399719, + 0.013581408187747002, + 0.02137528359889984, + 0.007241159211844206, + 0.006193045526742935, + 0.007632067427039146, + 0.004053245298564434, + 0.010598392225801945, + 0.011471089906990528, + 0.003741138381883502, + 0.006694066803902388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.05494910478591919 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.03809339925646782 + ], + "opposite_logprob_tokens": [ + -3.288093328475952 + ], + "logprob_reward": -0.03809339925646782, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809339925646782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04364091902971268, + 0.07075630128383636, + 0.03489441052079201, + 0.14104986190795898, + 0.042296357452869415, + 0.030452420935034752, + 0.26201632618904114, + 0.07439026236534119, + 0.275482714176178, + 0.025020405650138855 + ], + "credit_map": [ + 0.003994137980043888, + 0.006390376016497612, + 0.006312094163149595, + 0.006080516614019871, + 0.007435663603246212, + 0.006301836110651493, + 0.0034089505206793547, + 0.003717345418408513, + 0.002821745816618204, + 0.0118169654160738, + 0.0076300534419715405, + 0.011449338868260384, + 0.01345532014966011, + 0.016633732244372368, + 0.003315405687317252, + 0.003633741522207856, + 0.002800228539854288, + 0.007252216339111328, + 0.005313589237630367, + 0.0032498969230800867, + 0.005394517444074154, + 0.005514932330697775, + 0.0026857436168938875, + 0.0026832877192646265, + 0.01652498170733452, + 0.014454308897256851, + 0.01696009188890457, + 0.010481424629688263, + 0.029408132657408714, + 0.016600793227553368, + 0.010917272418737411, + 0.025702860206365585, + 0.006278680171817541, + 0.004754761699587107, + 0.006488021928817034, + 0.0041724336333572865, + 0.004181372467428446, + 0.006748638581484556, + 0.0038142327684909105, + 0.005858216434717178, + 0.003771264338865876, + 0.005485061090439558, + 0.005382138304412365, + 0.0027533741667866707, + 0.004020863212645054, + 0.0031337072141468525, + 0.003228261834010482, + 0.002677750075235963, + 0.05298482999205589, + 0.03819248825311661, + 0.030044499784708023, + 0.016986332833766937, + 0.04300576448440552, + 0.038007400929927826, + 0.02314659208059311, + 0.019648414105176926, + 0.00884538609534502, + 0.013522276654839516, + 0.010318138636648655, + 0.00865527056157589, + 0.010294601321220398, + 0.013237318955361843, + 0.0051116375252604485, + 0.004405631218105555, + 0.047012053430080414, + 0.07531841099262238, + 0.014020699076354504, + 0.011979293078184128, + 0.04936403036117554, + 0.056339848786592484, + 0.00744945602491498, + 0.013998933136463165, + 0.005170148331671953, + 0.0035504531115293503, + 0.0019022529013454914, + 0.001216671778820455, + 0.006353022065013647, + 0.0034625420812517405, + 0.0015602282946929336, + 0.0018050865037366748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.03809339925646782 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, full-body shot, angled slightly upward, captures a futuristic platform constructed of smooth white marble and polished metal, majestically suspended above a vast sea of clouds, against a backdrop of a deep twilight sky. A translucent train, composed entirely of countless tiny, shimmering geometric prisms, rests at the platform's edge. Its crystalline body emanates a soft, luminous white light from within, reflecting a rainbow of colors from every facet. Shot in hyperrealistic slow motion, the train slowly sets off with near-silence and silky smoothness, drifting leisurely to the right, clearing the platform's edge without any visible contact between wheels and track. As it moves, every pulse of light and diffraction of color is clearly amplified, its ever-changing brilliance reflected on the polished platform floor. The scene, rendered in exquisite detail, creates a visual spectacle of grandeur, tranquility, and profound dreaminess.", + "seed": 1011, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_1/index_3424_seed_1011_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": -0.02875549755990505, + "vlm_logprob": -0.02875549755990505, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02875549755990505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.024957256391644478 + ], + "opposite_logprob_tokens": [ + -4.024957180023193 + ], + "logprob_reward": -0.024957256391644478, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024957256391644478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3429309129714966, + 0.05405004322528839, + 0.04851444810628891, + 0.11260963976383209, + 0.032341714948415756, + 0.042813364416360855, + 0.2828451097011566, + 0.0338931530714035, + 0.023632947355508804, + 0.026368625462055206 + ], + "credit_map": [ + 0.03167768567800522, + 0.0466473326086998, + 0.017760733142495155, + 0.02697073109447956, + 0.0725010558962822, + 0.05684654042124748, + 0.04603026807308197, + 0.0444965735077858, + 0.006203943397849798, + 0.006344581954181194, + 0.003926044330000877, + 0.0038776551373302937, + 0.007084226235747337, + 0.011777304112911224, + 0.005490667652338743, + 0.009345624595880508, + 0.004189026076346636, + 0.00616843905299902, + 0.004041967447847128, + 0.0044148568995296955, + 0.008455712348222733, + 0.007120536640286446, + 0.008552621118724346, + 0.005571293644607067, + 0.006498425733298063, + 0.00951593928039074, + 0.00668909540399909, + 0.009793191216886044, + 0.01766449771821499, + 0.0157548189163208, + 0.027309179306030273, + 0.01938450336456299, + 0.004107789136469364, + 0.0036101785954087973, + 0.0030609765090048313, + 0.00349378096871078, + 0.005631496664136648, + 0.004080849699676037, + 0.005440993700176477, + 0.0029156540986150503, + 0.0038209406193345785, + 0.007768908049911261, + 0.0029043264221400023, + 0.004671321250498295, + 0.006717996671795845, + 0.008742054924368858, + 0.004101848229765892, + 0.004085970111191273, + 0.026787331327795982, + 0.08028735965490341, + 0.021201837807893753, + 0.024734044447541237, + 0.041274506598711014, + 0.03563106805086136, + 0.02602085843682289, + 0.02690812014043331, + 0.0026079437229782343, + 0.006631146650761366, + 0.002164009725674987, + 0.003544054925441742, + 0.005938755348324776, + 0.006433997303247452, + 0.002785115037113428, + 0.0037881354801356792, + 0.0023010147269815207, + 0.0029137940146028996, + 0.002022546250373125, + 0.0029041406232863665, + 0.003968921024352312, + 0.004170572385191917, + 0.0021219379268586636, + 0.0032300224993377924, + 0.003034866414964199, + 0.002449647057801485, + 0.0027815119829028845, + 0.0019235728541389108, + 0.006693079136312008, + 0.005584695842117071, + 0.0018917694687843323, + 0.0020094863139092922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.024957256391644478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.0059586502611637115 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.0059586502611637115, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059586502611637115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18100012838840485, + 0.110884889960289, + 0.10721906274557114, + 0.15035779774188995, + 0.023283906280994415, + 0.028048602864146233, + 0.19403131306171417, + 0.07520561665296555, + 0.08444414287805557, + 0.0455244816839695 + ], + "credit_map": [ + 0.015424521639943123, + 0.02175331301987171, + 0.015501933172345161, + 0.019987843930721283, + 0.025683127343654633, + 0.029209928587079048, + 0.028672827407717705, + 0.02476663887500763, + 0.011960177682340145, + 0.011300990357995033, + 0.008280728943645954, + 0.00998623576015234, + 0.011784052476286888, + 0.02354717068374157, + 0.011152664199471474, + 0.02287287451326847, + 0.013789678923785686, + 0.01709843799471855, + 0.009060376323759556, + 0.009222622029483318, + 0.013155871070921421, + 0.019666511565446854, + 0.013957076705992222, + 0.011268490925431252, + 0.009353586472570896, + 0.009553516283631325, + 0.008025850169360638, + 0.012980097904801369, + 0.017150528728961945, + 0.02430238388478756, + 0.04453371837735176, + 0.02445812150835991, + 0.0031750856433063745, + 0.0024726479314267635, + 0.0017855854239314795, + 0.0018927782075479627, + 0.0044806161895394325, + 0.0033738103229552507, + 0.0036189660895615816, + 0.0024844161234796047, + 0.0031132278963923454, + 0.005519921891391277, + 0.0019428094383329153, + 0.002712199231609702, + 0.003852908266708255, + 0.005187994800508022, + 0.0029333687853068113, + 0.0027861716225743294, + 0.017200052738189697, + 0.04524858295917511, + 0.02354930341243744, + 0.02098839357495308, + 0.0288360808044672, + 0.019957810640335083, + 0.020490366965532303, + 0.017760712653398514, + 0.005997770931571722, + 0.01377899944782257, + 0.004823730792850256, + 0.00904808845371008, + 0.014229916036128998, + 0.012116016820073128, + 0.008029879070818424, + 0.00718121649697423, + 0.007247758563607931, + 0.008320261724293232, + 0.005678216461092234, + 0.007191377691924572, + 0.022430209442973137, + 0.020433716475963593, + 0.00532723730430007, + 0.007815362885594368, + 0.004607192240655422, + 0.0052724373526871204, + 0.0032350735273212194, + 0.003000111086294055, + 0.011563889682292938, + 0.011384152807295322, + 0.002805267693474889, + 0.0036563551984727383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0059586502611637115 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.029798205941915512 + ], + "opposite_logprob_tokens": [ + -3.5297982692718506 + ], + "logprob_reward": -0.029798205941915512, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798205941915512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10259309411048889, + 0.20585300028324127, + 0.06409221142530441, + 0.09846749901771545, + 0.0477093867957592, + 0.0656430721282959, + 0.133150115609169, + 0.12623481452465057, + 0.06134661287069321, + 0.09491021931171417 + ], + "credit_map": [ + 0.010179790668189526, + 0.011813049204647541, + 0.010580511763691902, + 0.009913166053593159, + 0.02206447906792164, + 0.01811552420258522, + 0.009587655775249004, + 0.0103389210999012, + 0.019480254501104355, + 0.022276952862739563, + 0.01590808667242527, + 0.020482467487454414, + 0.029998507350683212, + 0.025116482749581337, + 0.040392227470874786, + 0.032198015600442886, + 0.0055982922203838825, + 0.009198679588735104, + 0.0049972692504525185, + 0.007962996140122414, + 0.006829824298620224, + 0.010160766541957855, + 0.011496891267597675, + 0.007847494445741177, + 0.007993895560503006, + 0.005080699920654297, + 0.008741472847759724, + 0.010967323556542397, + 0.010220203548669815, + 0.016080358996987343, + 0.026459574699401855, + 0.01292397454380989, + 0.0047106556594371796, + 0.005639700219035149, + 0.004410896450281143, + 0.004640444181859493, + 0.01010940968990326, + 0.007009184919297695, + 0.006061533000320196, + 0.005127565003931522, + 0.005301377270370722, + 0.014161919243633747, + 0.006631437223404646, + 0.007448884192854166, + 0.007510379422456026, + 0.006727932021021843, + 0.010858645662665367, + 0.007002492900937796, + 0.013763676397502422, + 0.0372847244143486, + 0.008735223673284054, + 0.01109231822192669, + 0.018606018275022507, + 0.016506807878613472, + 0.013970191590487957, + 0.013191147707402706, + 0.011078990995883942, + 0.012082523666322231, + 0.0066093397326767445, + 0.008047972805798054, + 0.034553077071905136, + 0.029732720926404, + 0.010547809302806854, + 0.01358236838132143, + 0.00607891334220767, + 0.007074147462844849, + 0.004660347010940313, + 0.006382819730788469, + 0.009474450722336769, + 0.009699576534330845, + 0.006845901720225811, + 0.011130459606647491, + 0.012011760845780373, + 0.009789659641683102, + 0.009881127625703812, + 0.007991600781679153, + 0.02156377211213112, + 0.019116448238492012, + 0.006209302227944136, + 0.008346542716026306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.029798205941915512 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.0789690613746643 + ], + "opposite_logprob_tokens": [ + -2.5789690017700195 + ], + "logprob_reward": -0.0789690613746643, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789690613746643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07661064714193344, + 0.07098246365785599, + 0.050542593002319336, + 0.05700499936938286, + 0.023206111043691635, + 0.09054376184940338, + 0.26768481731414795, + 0.06407615542411804, + 0.07358263432979584, + 0.22576580941677094 + ], + "credit_map": [ + 0.007923471741378307, + 0.0117839016020298, + 0.005910174455493689, + 0.008934997022151947, + 0.014120301231741905, + 0.010316888801753521, + 0.00910578016191721, + 0.008515129797160625, + 0.006398470606654882, + 0.006502329837530851, + 0.005834002047777176, + 0.005781868938356638, + 0.00756203755736351, + 0.010202775709331036, + 0.011825782246887684, + 0.01687519997358322, + 0.004434310365468264, + 0.007038856856524944, + 0.005219672806560993, + 0.004774074535816908, + 0.0064411344937980175, + 0.00770777091383934, + 0.005791807547211647, + 0.009134965017437935, + 0.004097494296729565, + 0.005345821380615234, + 0.005224580876529217, + 0.006119930185377598, + 0.008958298712968826, + 0.007176402024924755, + 0.011022993363440037, + 0.009059479460120201, + 0.002923813182860613, + 0.002898477017879486, + 0.0015168730169534683, + 0.0014822585508227348, + 0.00586694898083806, + 0.0035888534039258957, + 0.002276382641866803, + 0.0026525063440203667, + 0.00789256114512682, + 0.023200124502182007, + 0.005588287953287363, + 0.010133599862456322, + 0.020190317183732986, + 0.009595316834747791, + 0.008363665081560612, + 0.0055798934772610664, + 0.024151824414730072, + 0.031350016593933105, + 0.04111967235803604, + 0.029899924993515015, + 0.04168891906738281, + 0.02658257633447647, + 0.03801518678665161, + 0.034876689314842224, + 0.005733688827604055, + 0.01055190246552229, + 0.003849869593977928, + 0.004143367055803537, + 0.016717692837119102, + 0.01294032484292984, + 0.004948989953845739, + 0.005190317519009113, + 0.0049741147086024284, + 0.005747657269239426, + 0.0058892155066132545, + 0.005371395964175463, + 0.019680406898260117, + 0.01859736442565918, + 0.005541635677218437, + 0.007780841551721096, + 0.0501164086163044, + 0.03063303604722023, + 0.020779883489012718, + 0.013436369597911835, + 0.04088909178972244, + 0.032025229185819626, + 0.015371153131127357, + 0.022514639422297478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.0789690613746643 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a translucent train composed of tiny shimmering geometric prisms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train emitting a soft luminous white light that reflects rainbow colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set on a futuristic platform made of smooth white marble and polished metal suspended above clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the train drift leisurely to the right while clearing the platform's edge without visible contact between wheels and track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady full-body view angled slightly upward against a deep twilight 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": [ + -0.004094313830137253 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094313830137253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094313830137253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06563258916139603, + 0.04573503136634827, + 0.04799456521868706, + 0.04280467331409454, + 0.02406192757189274, + 0.12413706630468369, + 0.2829577624797821, + 0.03821907564997673, + 0.12274156510829926, + 0.20571574568748474 + ], + "credit_map": [ + 0.008140422403812408, + 0.008411522954702377, + 0.0068250312469899654, + 0.005910332780331373, + 0.013053418137133121, + 0.009386097081005573, + 0.006002617068588734, + 0.0079031428322196, + 0.004200059454888105, + 0.007942981086671352, + 0.00413033040240407, + 0.005016131792217493, + 0.005732967518270016, + 0.005803371779620647, + 0.005335748661309481, + 0.0075734397396445274, + 0.0037425560876727104, + 0.006584479007869959, + 0.004525030963122845, + 0.004835022613406181, + 0.008954420685768127, + 0.00771859148517251, + 0.00557892257347703, + 0.006055541802197695, + 0.0058595603331923485, + 0.004797694738954306, + 0.007267627865076065, + 0.004660444799810648, + 0.007329544518142939, + 0.005178287625312805, + 0.003971056547015905, + 0.003740456886589527, + 0.0028052509296685457, + 0.0026090526953339577, + 0.0026177475228905678, + 0.0022813219111412764, + 0.003854550654068589, + 0.003601514268666506, + 0.0026466010604053736, + 0.003645888762548566, + 0.01025929395109415, + 0.03150438144803047, + 0.00883428193628788, + 0.010617786087095737, + 0.028234966099262238, + 0.009839575737714767, + 0.013922329992055893, + 0.010924448259174824, + 0.024968339130282402, + 0.08935442566871643, + 0.02274240180850029, + 0.025779390707612038, + 0.041188906878232956, + 0.03307913988828659, + 0.020797226577997208, + 0.025047939270734787, + 0.00530345318838954, + 0.007020116783678532, + 0.002826125593855977, + 0.0028827288188040257, + 0.006336118094623089, + 0.006286958232522011, + 0.0030983262695372105, + 0.0044652498327195644, + 0.01253036130219698, + 0.014566043391823769, + 0.007124710362404585, + 0.011689330451190472, + 0.027152108028531075, + 0.025683462619781494, + 0.009732494130730629, + 0.014263051562011242, + 0.021009331569075584, + 0.01606052555143833, + 0.01815500482916832, + 0.013221241533756256, + 0.06210239231586456, + 0.05256292596459389, + 0.009024253115057945, + 0.013580072671175003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.004094313830137253 + } + ], + "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/0000110/group_10/index_tensor([6748, 6748])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_tensor([6748, 6748])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..72488db1fa3daa2e6cf1e19a6635927b62bf9064 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_tensor([6748, 6748])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a dynamic shot reminiscent of Spider-Man: Into the Spider-Verse, the camera zooms forward at breakneck speed, capturing an absurd close-up. The subject is a cartoon character with a distinctly clay-like texture, its surface bearing subtle finger impressions, imbued with a handmade quality. Clad in a vibrant, saturated yellow jumpsuit, the figure tilts its head back, eyes wide open, an expression of innocent, focused anticipation. It is set against a backdrop woven from surreal pop art and a constantly flickering glitch aesthetic. Large, flatly painted geometric shapes interweave, rip, and reassemble with comic-book-like Ben-Day dots, filling the space with shifted chromatic aberrations and pixelated digital interference. The character holds aloft a cylindrical can of potato chips, the label clearly inscribed with \"Yogurt Flavor\" in bold green font. He decisively inverts the can, directly above his head. With one steady movement, the can's entire contents, complete with saddle-shaped chips, cascade down like an eerie waterfall, accompanied by a fine burst of green seasoning powder, instantly covering his entire head and face. The rapidly zoomed-in camera captures the shattering of each chip and the cascading of dust, blending seamlessly with the background's visual noise, ultimately framing the character's comical face after being \"cleaned\" by the chips, creating a chaotic and visually striking atmosphere of psychedelic comedy.", + "seed": 22275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_6748_seed_22275_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": -0.014643462048843503, + "vlm_logprob": -0.014643462048843503, + "vlm_yes_no_margin": 4.649999952316284, + "vlm_token_credit": -0.014643462048843503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02014346607029438 + ], + "opposite_logprob_tokens": [ + -5.145143508911133 + ], + "logprob_reward": -0.02014346607029438, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.02014346607029438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11746378988027573, + 0.03349652513861656, + 0.07865205407142639, + 0.0747823417186737, + 0.15850602090358734, + 0.03526321426033974, + 0.12724019587039948, + 0.14343421161174774, + 0.10343845188617706, + 0.12772323191165924 + ], + "credit_map": [ + 0.016573140397667885, + 0.007980822585523129, + 0.016776399686932564, + 0.013176200911402702, + 0.01791982538998127, + 0.008231683634221554, + 0.018529875203967094, + 0.018275832757353783, + 0.005726230796426535, + 0.003445265581831336, + 0.0037603278178721666, + 0.004715489223599434, + 0.004287524148821831, + 0.0036122403107583523, + 0.004250905942171812, + 0.0036985392216593027, + 0.007179224863648415, + 0.007865538820624352, + 0.006193501874804497, + 0.009568891488015652, + 0.015213980339467525, + 0.00870450772345066, + 0.014970526099205017, + 0.00895589031279087, + 0.006268948782235384, + 0.007630384061485529, + 0.008981749415397644, + 0.0077384887263178825, + 0.016488658264279366, + 0.009430881589651108, + 0.010336090810596943, + 0.007907139137387276, + 0.012251093052327633, + 0.02396811917424202, + 0.007738680113106966, + 0.02778255194425583, + 0.00995871052145958, + 0.007970803417265415, + 0.05945894867181778, + 0.009377111680805683, + 0.0033481295686215162, + 0.0031457950826734304, + 0.0050710090436041355, + 0.0035459273494780064, + 0.006915612146258354, + 0.00340472930110991, + 0.0061858766712248325, + 0.003646132303401828, + 0.0098721357062459, + 0.010063991881906986, + 0.017388666048645973, + 0.01749521493911743, + 0.023984363302588463, + 0.008210841566324234, + 0.028946181759238243, + 0.011278785765171051, + 0.02776680886745453, + 0.01622549630701542, + 0.01710451766848564, + 0.019935665652155876, + 0.015190932899713516, + 0.012767625972628593, + 0.02126634679734707, + 0.013176816515624523, + 0.012792137451469898, + 0.01489852461963892, + 0.012375524267554283, + 0.01408497616648674, + 0.01660403423011303, + 0.010796823538839817, + 0.012941570021212101, + 0.008944867178797722, + 0.007345203775912523, + 0.008391211740672588, + 0.03900359198451042, + 0.02051294595003128, + 0.011299106292426586, + 0.005705670453608036, + 0.024858783930540085, + 0.010606714524328709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.02014346607029438 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011102918535470963 + ], + "opposite_logprob_tokens": [ + -4.51110315322876 + ], + "logprob_reward": -0.011102918535470963, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011102918535470963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23275084793567657, + 0.07812068611383438, + 0.02738417312502861, + 0.08312669396400452, + 0.02890373207628727, + 0.0746183693408966, + 0.03610147908329964, + 0.18452276289463043, + 0.12243135273456573, + 0.1320398598909378 + ], + "credit_map": [ + 0.022761236876249313, + 0.017364874482154846, + 0.02544499561190605, + 0.0291883684694767, + 0.03898040950298309, + 0.02032569795846939, + 0.027247058227658272, + 0.05143820866942406, + 0.009426020085811615, + 0.006150306202471256, + 0.005990653298795223, + 0.014153366908431053, + 0.015588164329528809, + 0.008044595830142498, + 0.009397969581186771, + 0.009369607083499432, + 0.0028831164818257093, + 0.0024315090849995613, + 0.0039597162976861, + 0.0037744478322565556, + 0.0033555831760168076, + 0.003160343272611499, + 0.0038367770612239838, + 0.003982679918408394, + 0.006043685134500265, + 0.008211719803512096, + 0.00768031133338809, + 0.017189117148518562, + 0.01171915978193283, + 0.0067140208557248116, + 0.012629848904907703, + 0.012938828207552433, + 0.002619346836581826, + 0.004226157907396555, + 0.004088935907930136, + 0.0036409839522093534, + 0.004585815127938986, + 0.0025466473307460546, + 0.004530150908976793, + 0.0026656941045075655, + 0.005866963881999254, + 0.01204829290509224, + 0.009775541722774506, + 0.013346973806619644, + 0.009916764684021473, + 0.004494440276175737, + 0.011884210631251335, + 0.0072851767763495445, + 0.00434023467823863, + 0.003627820871770382, + 0.004568715579807758, + 0.0049032592214643955, + 0.005941387265920639, + 0.002315659075975418, + 0.006679003592580557, + 0.003725397167727351, + 0.023398401215672493, + 0.023444846272468567, + 0.01768646389245987, + 0.02171557955443859, + 0.02090570516884327, + 0.032814983278512955, + 0.02376118302345276, + 0.02079560048878193, + 0.018776917830109596, + 0.007147816475480795, + 0.0186692513525486, + 0.02105727232992649, + 0.013190110214054585, + 0.00942957028746605, + 0.01843235082924366, + 0.0157280582934618, + 0.00796687975525856, + 0.013981125317513943, + 0.024231191724538803, + 0.03528870269656181, + 0.016240891069173813, + 0.01191037893295288, + 0.014153074473142624, + 0.008267627097666264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.011102918535470963 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202053382992744 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.018202053382992744, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018202053382992744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2561643123626709, + 0.13558240234851837, + 0.10785113275051117, + 0.039156723767519, + 0.04995700716972351, + 0.02095109410583973, + 0.06134403869509697, + 0.06864448636770248, + 0.048316244035959244, + 0.21203261613845825 + ], + "credit_map": [ + 0.03985977917909622, + 0.015611069276928902, + 0.025658080354332924, + 0.03699182718992233, + 0.028012501075863838, + 0.017062639817595482, + 0.05922280624508858, + 0.033745601773262024, + 0.023800447583198547, + 0.01310683973133564, + 0.019897686317563057, + 0.014651316218078136, + 0.019699839875102043, + 0.012226024642586708, + 0.018554778769612312, + 0.0136454738676548, + 0.0054314169101417065, + 0.00889070425182581, + 0.009553788229823112, + 0.014270463027060032, + 0.021174240857362747, + 0.011993931606411934, + 0.017137864604592323, + 0.019398730248212814, + 0.0032265176996588707, + 0.0028928881511092186, + 0.003561622928828001, + 0.004420132841914892, + 0.00867859274148941, + 0.004009569995105267, + 0.007308547850698233, + 0.005058850161731243, + 0.00376979261636734, + 0.004822758492082357, + 0.0037349385675042868, + 0.010416348464787006, + 0.003875162685289979, + 0.0026218006387352943, + 0.017080888152122498, + 0.003635317785665393, + 0.002228126861155033, + 0.0014609895879402757, + 0.0016791446832939982, + 0.0030189668759703636, + 0.001971908612176776, + 0.0016516281757503748, + 0.006688074674457312, + 0.002252254169434309, + 0.007937196642160416, + 0.003962208982557058, + 0.016077036038041115, + 0.008373100310564041, + 0.005664934404194355, + 0.0040174429304897785, + 0.009960952214896679, + 0.005351165775209665, + 0.01042002160102129, + 0.007068722974509001, + 0.006845541298389435, + 0.0071669952012598515, + 0.007654974702745676, + 0.012097777798771858, + 0.008932671509683132, + 0.008457778953015804, + 0.005713092628866434, + 0.004559366498142481, + 0.0062307557091116905, + 0.006306529976427555, + 0.00872571486979723, + 0.0050190226174890995, + 0.006200168281793594, + 0.00556159345433116, + 0.011246695183217525, + 0.017042335122823715, + 0.03726036101579666, + 0.08062247931957245, + 0.0203627310693264, + 0.005979018285870552, + 0.03300650417804718, + 0.006512496154755354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.018202053382992744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031966085080057383 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.0031966085080057383, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031966085080057383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13581785559654236, + 0.037591345608234406, + 0.05395698547363281, + 0.08085580170154572, + 0.027399184182286263, + 0.023141926154494286, + 0.07022803276777267, + 0.07211946696043015, + 0.04678155854344368, + 0.45210787653923035 + ], + "credit_map": [ + 0.024048399180173874, + 0.008811065927147865, + 0.015509717166423798, + 0.01877862587571144, + 0.01858169585466385, + 0.007930751889944077, + 0.028353113681077957, + 0.01380447018891573, + 0.0068234666250646114, + 0.0036017862148582935, + 0.005476903170347214, + 0.0038973307237029076, + 0.004577250219881535, + 0.0036086863838136196, + 0.0049300468526780605, + 0.004675873089581728, + 0.0038077901117503643, + 0.00561104342341423, + 0.004162698518484831, + 0.007173569872975349, + 0.01082516834139824, + 0.004103883635252714, + 0.008860229514539242, + 0.009412596002221107, + 0.0063189235515892506, + 0.011430067010223866, + 0.005788162816315889, + 0.012616407126188278, + 0.008265489712357521, + 0.006997679825872183, + 0.02124887704849243, + 0.008190188556909561, + 0.0022728382609784603, + 0.0036015824880450964, + 0.003130246652290225, + 0.005718751810491085, + 0.0033914141822606325, + 0.001498651341535151, + 0.006433479953557253, + 0.0013522182125598192, + 0.002570958575233817, + 0.002117420081049204, + 0.003496144199743867, + 0.0033704484812915325, + 0.00296471593901515, + 0.0016492020804435015, + 0.004883601795881987, + 0.002089434303343296, + 0.00767207145690918, + 0.007550645619630814, + 0.01107813511043787, + 0.009006270207464695, + 0.01230099517852068, + 0.005546593107283115, + 0.011073945090174675, + 0.005999376997351646, + 0.01217618491500616, + 0.008487158454954624, + 0.0076306830160319805, + 0.006186455022543669, + 0.00818873941898346, + 0.01422608457505703, + 0.008527928963303566, + 0.006696230731904507, + 0.008871941827237606, + 0.0031802922021597624, + 0.00653160410001874, + 0.00666407123208046, + 0.006274076644331217, + 0.004614308010786772, + 0.00587861193343997, + 0.004766649566590786, + 0.020112916827201843, + 0.04943164065480232, + 0.05816582217812538, + 0.13953113555908203, + 0.05886605754494667, + 0.026980085298419, + 0.07826712727546692, + 0.020753039047122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0031966085080057383 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057226374745369 + ], + "opposite_logprob_tokens": [ + -3.8955721855163574 + ], + "logprob_reward": -0.02057226374745369, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057226374745369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1715056151151657, + 0.04108972102403641, + 0.053314872086048126, + 0.19926497340202332, + 0.10912586003541946, + 0.015249178744852543, + 0.03699371591210365, + 0.09256014227867126, + 0.15549197793006897, + 0.1254040002822876 + ], + "credit_map": [ + 0.031709086149930954, + 0.009789266623556614, + 0.022123008966445923, + 0.021422555670142174, + 0.017680983990430832, + 0.009989290498197079, + 0.035321757197380066, + 0.023469651117920876, + 0.007750825956463814, + 0.0034512565471231937, + 0.006629765499383211, + 0.006066709291189909, + 0.0047014690935611725, + 0.0024980069138109684, + 0.006269071251153946, + 0.003722613211721182, + 0.005812123883515596, + 0.00404534162953496, + 0.006999094970524311, + 0.009800593368709087, + 0.0045048147439956665, + 0.00865029264241457, + 0.005991630721837282, + 0.007510976865887642, + 0.017149507999420166, + 0.019630558788776398, + 0.022791247814893723, + 0.028202831745147705, + 0.030073557049036026, + 0.020627884194254875, + 0.030376482754945755, + 0.03041289933025837, + 0.007767748553305864, + 0.01734095625579357, + 0.007363052107393742, + 0.01711675524711609, + 0.00742069398984313, + 0.005404424853622913, + 0.039474520832300186, + 0.007237713783979416, + 0.0013077177572995424, + 0.001639082096517086, + 0.0036193884443491697, + 0.001977139851078391, + 0.0017326812958344817, + 0.0011021216632798314, + 0.0024120898451656103, + 0.0014589570928364992, + 0.004614592529833317, + 0.0027010610792785883, + 0.011810063384473324, + 0.0034803017042577267, + 0.003697044914588332, + 0.002023134147748351, + 0.005825355648994446, + 0.002842160640284419, + 0.016352534294128418, + 0.008210066705942154, + 0.009308287873864174, + 0.013439148664474487, + 0.008122866973280907, + 0.013379975222051144, + 0.014627787284553051, + 0.009119473397731781, + 0.01936349831521511, + 0.008764752186834812, + 0.022128552198410034, + 0.025163380429148674, + 0.025351108983159065, + 0.014068695716559887, + 0.026607459411025047, + 0.014044531621038914, + 0.005894279573112726, + 0.006457868032157421, + 0.03235546499490738, + 0.050115421414375305, + 0.008340174332261086, + 0.002920699305832386, + 0.01636786200106144, + 0.0029522343538701534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.02057226374745369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dynamic shot reminiscent of Spider-Man: Into the Spider-Verse, the camera zooms forward at breakneck speed, capturing an absurd close-up. The subject is a cartoon character with a distinctly clay-like texture, its surface bearing subtle finger impressions, imbued with a handmade quality. Clad in a vibrant, saturated yellow jumpsuit, the figure tilts its head back, eyes wide open, an expression of innocent, focused anticipation. It is set against a backdrop woven from surreal pop art and a constantly flickering glitch aesthetic. Large, flatly painted geometric shapes interweave, rip, and reassemble with comic-book-like Ben-Day dots, filling the space with shifted chromatic aberrations and pixelated digital interference. The character holds aloft a cylindrical can of potato chips, the label clearly inscribed with \"Yogurt Flavor\" in bold green font. He decisively inverts the can, directly above his head. With one steady movement, the can's entire contents, complete with saddle-shaped chips, cascade down like an eerie waterfall, accompanied by a fine burst of green seasoning powder, instantly covering his entire head and face. The rapidly zoomed-in camera captures the shattering of each chip and the cascading of dust, blending seamlessly with the background's visual noise, ultimately framing the character's comical face after being \"cleaned\" by the chips, creating a chaotic and visually striking atmosphere of psychedelic comedy.", + "seed": 22276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_6748_seed_22276_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": -0.013291808590292931, + "vlm_logprob": -0.013291808590292931, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.013291808590292931, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017280373722314835 + ], + "opposite_logprob_tokens": [ + -4.767280578613281 + ], + "logprob_reward": -0.017280373722314835, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017280373722314835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21274498105049133, + 0.07748415321111679, + 0.1184324100613594, + 0.06454171985387802, + 0.03834058716893196, + 0.05972357839345932, + 0.0959279015660286, + 0.18020695447921753, + 0.08092612773180008, + 0.07167161256074905 + ], + "credit_map": [ + 0.028630699962377548, + 0.024113375693559647, + 0.022570425644516945, + 0.02480538934469223, + 0.021856190636754036, + 0.024279111996293068, + 0.022127658128738403, + 0.044362131506204605, + 0.00712076248601079, + 0.006700599100440741, + 0.007484711706638336, + 0.010686645284295082, + 0.014615644700825214, + 0.008111792616546154, + 0.012680251151323318, + 0.01008374523371458, + 0.009271757677197456, + 0.020527733489871025, + 0.009090214036405087, + 0.011328148655593395, + 0.011431748047471046, + 0.008489140309393406, + 0.0249541737139225, + 0.023339496925473213, + 0.006120521109551191, + 0.005293745081871748, + 0.005377947818487883, + 0.011868979781866074, + 0.007132112514227629, + 0.00879292469471693, + 0.011355374939739704, + 0.008600112982094288, + 0.00761163467541337, + 0.004999951925128698, + 0.006463987752795219, + 0.003875137073919177, + 0.002715067006647587, + 0.0038051800802350044, + 0.003995947539806366, + 0.004873679950833321, + 0.005848279222846031, + 0.008033559657633305, + 0.011578601785004139, + 0.00788410846143961, + 0.0069953678175807, + 0.006096165161579847, + 0.0063665504567325115, + 0.006920950952917337, + 0.010570465587079525, + 0.0077818031422793865, + 0.02001940831542015, + 0.01519391406327486, + 0.010321419686079025, + 0.007543283049017191, + 0.010757378302514553, + 0.013740229420363903, + 0.012683655135333538, + 0.01784588024020195, + 0.011258252896368504, + 0.050617512315511703, + 0.014405746944248676, + 0.02290898934006691, + 0.02537035010755062, + 0.0251165758818388, + 0.012074640020728111, + 0.007007092237472534, + 0.006467584054917097, + 0.009579210542142391, + 0.008014515973627567, + 0.008070170879364014, + 0.01637200266122818, + 0.013340908102691174, + 0.010065602138638496, + 0.006986778695136309, + 0.011489484459161758, + 0.011535082943737507, + 0.005747381132096052, + 0.004728695377707481, + 0.011960010975599289, + 0.009158574976027012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.017280373722314835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650598883628845 + ], + "opposite_logprob_tokens": [ + -4.882650375366211 + ], + "logprob_reward": -0.007650598883628845, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650598883628845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14391079545021057, + 0.16436214745044708, + 0.025476284325122833, + 0.14180676639080048, + 0.02279210276901722, + 0.12228094786405563, + 0.05338394641876221, + 0.07117612659931183, + 0.11966623365879059, + 0.1351446509361267 + ], + "credit_map": [ + 0.019793406128883362, + 0.01443255040794611, + 0.016086187213659286, + 0.016429755836725235, + 0.022266723215579987, + 0.015989335253834724, + 0.015110218897461891, + 0.02380261942744255, + 0.011141173541545868, + 0.013141988776624203, + 0.009886862710118294, + 0.017127925530076027, + 0.05763022601604462, + 0.019705073907971382, + 0.023467516526579857, + 0.012261386029422283, + 0.0020475038327276707, + 0.0028992779552936554, + 0.002904294990003109, + 0.003590121166780591, + 0.0031829909421503544, + 0.0034542290959507227, + 0.003440810600295663, + 0.003957053646445274, + 0.010504212230443954, + 0.008419849909842014, + 0.009271676652133465, + 0.02037058398127556, + 0.015486498363316059, + 0.021278781816363335, + 0.02813921868801117, + 0.028335953131318092, + 0.003950128797441721, + 0.0024197129532694817, + 0.0033178923185914755, + 0.001770075410604477, + 0.002648055786266923, + 0.0023487149737775326, + 0.002683904953300953, + 0.0036536171101033688, + 0.014318953268229961, + 0.014422296546399593, + 0.02979006990790367, + 0.011392325162887573, + 0.013229715637862682, + 0.013660003431141376, + 0.011810284107923508, + 0.013657302595674992, + 0.00599035806953907, + 0.004896167200058699, + 0.01239326223731041, + 0.005902641452848911, + 0.004445151425898075, + 0.003850580658763647, + 0.008454812690615654, + 0.007450971752405167, + 0.004418211057782173, + 0.008068729192018509, + 0.005198238883167505, + 0.0223793126642704, + 0.0049695782363414764, + 0.00721839489415288, + 0.011020170524716377, + 0.007903490215539932, + 0.00706529151648283, + 0.010648497380316257, + 0.006132402457296848, + 0.016702381893992424, + 0.010890210047364235, + 0.0056281001307070255, + 0.03828052803874016, + 0.02431883104145527, + 0.008464327082037926, + 0.01987002231180668, + 0.00872341450303793, + 0.02051457203924656, + 0.019198354333639145, + 0.007307498715817928, + 0.03178542107343674, + 0.019281035289168358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.007650598883628845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357226073741913 + ], + "opposite_logprob_tokens": [ + -3.6513571739196777 + ], + "logprob_reward": -0.026357226073741913, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357226073741913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09610722959041595, + 0.3519383370876312, + 0.07160157710313797, + 0.05928787589073181, + 0.03634343668818474, + 0.05163517966866493, + 0.06875214725732803, + 0.06905417144298553, + 0.07229509204626083, + 0.12298498302698135 + ], + "credit_map": [ + 0.008937797509133816, + 0.009887241758406162, + 0.007007567677646875, + 0.017673585563898087, + 0.017907774075865746, + 0.009111212566494942, + 0.011941041797399521, + 0.013641012832522392, + 0.0364019013941288, + 0.06786862760782242, + 0.022344011813402176, + 0.03318500518798828, + 0.06344326585531235, + 0.020886793732643127, + 0.08304465562105179, + 0.024764062836766243, + 0.005253801587969065, + 0.011040483601391315, + 0.006860785186290741, + 0.008294347673654556, + 0.009417936205863953, + 0.012477066367864609, + 0.00862801168113947, + 0.00962914526462555, + 0.008114602416753769, + 0.005787484813481569, + 0.005547236651182175, + 0.005747874267399311, + 0.005579076241701841, + 0.007327781058847904, + 0.009630678221583366, + 0.011553140357136726, + 0.008058895356953144, + 0.004541210364550352, + 0.004853993188589811, + 0.003682067384943366, + 0.003983831498771906, + 0.0030480530112981796, + 0.003749622730538249, + 0.0044257608242332935, + 0.005767153110355139, + 0.006094127427786589, + 0.005470666103065014, + 0.005936814472079277, + 0.008122247643768787, + 0.008225413039326668, + 0.0058896783739328384, + 0.006129076704382896, + 0.007590698078274727, + 0.00614382466301322, + 0.009349312633275986, + 0.010120702907443047, + 0.010211098939180374, + 0.00645294226706028, + 0.009630639106035233, + 0.009252931922674179, + 0.009286007843911648, + 0.00891108624637127, + 0.006467720959335566, + 0.010397935286164284, + 0.008265757001936436, + 0.007301941979676485, + 0.0077258991077542305, + 0.010697823017835617, + 0.0064554400742053986, + 0.006987494882196188, + 0.005453480873256922, + 0.012450458481907845, + 0.005722230765968561, + 0.0065768565982580185, + 0.015455464832484722, + 0.013193663209676743, + 0.009002250619232655, + 0.016045674681663513, + 0.008491993881762028, + 0.020499534904956818, + 0.015982385724782944, + 0.005794857162982225, + 0.028940174728631973, + 0.018228115513920784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.026357226073741913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06469561159610748, + 0.18976733088493347, + 0.05278201773762703, + 0.09335944056510925, + 0.05674232542514801, + 0.09050940722227097, + 0.13430850207805634, + 0.14556115865707397, + 0.07129579782485962, + 0.10097838938236237 + ], + "credit_map": [ + 0.0111369164660573, + 0.011084177531301975, + 0.004284059163182974, + 0.011013609357178211, + 0.006286529824137688, + 0.005498407408595085, + 0.007277003023773432, + 0.008114917203783989, + 0.019460318610072136, + 0.020390819758176804, + 0.01031000167131424, + 0.016402777284383774, + 0.02394501119852066, + 0.036081355065107346, + 0.047858767211437225, + 0.015318290330469608, + 0.004063165280967951, + 0.006965238135308027, + 0.005195398814976215, + 0.006928838323801756, + 0.007962421514093876, + 0.0084411995485425, + 0.006874407175928354, + 0.0063513536006212234, + 0.010406355373561382, + 0.006321529392153025, + 0.008461968041956425, + 0.011869234032928944, + 0.0071761272847652435, + 0.007236461620777845, + 0.02024831995368004, + 0.021639447659254074, + 0.007010639179497957, + 0.006267269141972065, + 0.00800557341426611, + 0.005016941577196121, + 0.011277326382696629, + 0.005102161318063736, + 0.006993105169385672, + 0.007069316226989031, + 0.008614224381744862, + 0.012897709384560585, + 0.015753090381622314, + 0.012210482731461525, + 0.012277130968868732, + 0.00881276186555624, + 0.010569042526185513, + 0.009374971501529217, + 0.013757456094026566, + 0.00909379031509161, + 0.031359754502773285, + 0.017941972240805626, + 0.013252127915620804, + 0.010800009593367577, + 0.01679706759750843, + 0.021306343376636505, + 0.015728659927845, + 0.016795895993709564, + 0.011744575574994087, + 0.03063581883907318, + 0.012632634490728378, + 0.015626458451151848, + 0.020028548315167427, + 0.022368580102920532, + 0.00535985454916954, + 0.0053268056362867355, + 0.009558392688632011, + 0.008845491334795952, + 0.005258176475763321, + 0.005368323996663094, + 0.013871178962290287, + 0.0177075807005167, + 0.009655723348259926, + 0.013190374709665775, + 0.011303916573524475, + 0.013957899063825607, + 0.012630256824195385, + 0.0068809883669018745, + 0.019097870215773582, + 0.014261369593441486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.004097281955182552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073562316596508 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073562316596508, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073562316596508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0976271778345108, + 0.1624777764081955, + 0.14621661603450775, + 0.08176436275243759, + 0.061761703342199326, + 0.05775391682982445, + 0.15656568109989166, + 0.04239026457071304, + 0.11590468883514404, + 0.0775378867983818 + ], + "credit_map": [ + 0.018427662551403046, + 0.012532680295407772, + 0.007769786287099123, + 0.010911442339420319, + 0.012641270644962788, + 0.008215435780584812, + 0.009067102335393429, + 0.018061796203255653, + 0.01620797999203205, + 0.01682574674487114, + 0.016616784036159515, + 0.014332986436784267, + 0.03531933203339577, + 0.0342874601483345, + 0.015459207817912102, + 0.013428271748125553, + 0.015042954124510288, + 0.03887908533215523, + 0.013439660891890526, + 0.0175416748970747, + 0.015930337831377983, + 0.013237782754004002, + 0.019704017788171768, + 0.012441095896065235, + 0.007408750243484974, + 0.008300172165036201, + 0.00961231254041195, + 0.011026797816157341, + 0.010987168177962303, + 0.013554275967180729, + 0.011686357669532299, + 0.009188530966639519, + 0.010978474281728268, + 0.006629811599850655, + 0.011971534229815006, + 0.006057532504200935, + 0.004744559992104769, + 0.005831892602145672, + 0.0068165697157382965, + 0.008731319569051266, + 0.008203433826565742, + 0.005207863170653582, + 0.01343284547328949, + 0.006139224395155907, + 0.00628763297572732, + 0.005864775273948908, + 0.006536693777889013, + 0.006081441882997751, + 0.02003483846783638, + 0.01351456344127655, + 0.03159269690513611, + 0.01573740318417549, + 0.016392771154642105, + 0.013617211952805519, + 0.023673484101891518, + 0.022002704441547394, + 0.004018779378384352, + 0.008482363075017929, + 0.0028875390999019146, + 0.00788204651325941, + 0.0048576947301626205, + 0.00403093546628952, + 0.0054373848251998425, + 0.004793517757207155, + 0.010091842152178288, + 0.011861475184559822, + 0.008954987861216068, + 0.012582344934344292, + 0.010847818106412888, + 0.008239621296525002, + 0.03301825001835823, + 0.02030833624303341, + 0.007047583814710379, + 0.009760619141161442, + 0.0070269424468278885, + 0.011576615273952484, + 0.009965207427740097, + 0.0047449697740375996, + 0.01482909545302391, + 0.012586846947669983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.011073562316596508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dynamic shot reminiscent of Spider-Man: Into the Spider-Verse, the camera zooms forward at breakneck speed, capturing an absurd close-up. The subject is a cartoon character with a distinctly clay-like texture, its surface bearing subtle finger impressions, imbued with a handmade quality. Clad in a vibrant, saturated yellow jumpsuit, the figure tilts its head back, eyes wide open, an expression of innocent, focused anticipation. It is set against a backdrop woven from surreal pop art and a constantly flickering glitch aesthetic. Large, flatly painted geometric shapes interweave, rip, and reassemble with comic-book-like Ben-Day dots, filling the space with shifted chromatic aberrations and pixelated digital interference. The character holds aloft a cylindrical can of potato chips, the label clearly inscribed with \"Yogurt Flavor\" in bold green font. He decisively inverts the can, directly above his head. With one steady movement, the can's entire contents, complete with saddle-shaped chips, cascade down like an eerie waterfall, accompanied by a fine burst of green seasoning powder, instantly covering his entire head and face. The rapidly zoomed-in camera captures the shattering of each chip and the cascading of dust, blending seamlessly with the background's visual noise, ultimately framing the character's comical face after being \"cleaned\" by the chips, creating a chaotic and visually striking atmosphere of psychedelic comedy.", + "seed": 22277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_6748_seed_22277_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": -0.02025228748098016, + "vlm_logprob": -0.02025228748098016, + "vlm_yes_no_margin": 4.124999952316284, + "vlm_token_credit": -0.02025228748098016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026818417012691498 + ], + "opposite_logprob_tokens": [ + -4.02681827545166 + ], + "logprob_reward": -0.026818417012691498, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.026818417012691498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1559603363275528, + 0.20592765510082245, + 0.06570979952812195, + 0.03434295207262039, + 0.09098177403211594, + 0.0858304426074028, + 0.09040859341621399, + 0.14929240942001343, + 0.04013149067759514, + 0.08141462504863739 + ], + "credit_map": [ + 0.013226855546236038, + 0.017557358369231224, + 0.01636943407356739, + 0.03924866020679474, + 0.02051299810409546, + 0.014918205328285694, + 0.0195339135825634, + 0.014592884108424187, + 0.020964104682207108, + 0.019942890852689743, + 0.01653667539358139, + 0.02028312347829342, + 0.0433661974966526, + 0.0259512048214674, + 0.029431594535708427, + 0.029451854526996613, + 0.0103182103484869, + 0.006420282647013664, + 0.009432094171643257, + 0.007917448878288269, + 0.009247600100934505, + 0.004727411549538374, + 0.008614327758550644, + 0.0090324142947793, + 0.0044488124549388885, + 0.00465551158413291, + 0.003282008459791541, + 0.0026629306375980377, + 0.008060985244810581, + 0.0037689199671149254, + 0.004215438850224018, + 0.003248340217396617, + 0.0062910630367696285, + 0.010182587429881096, + 0.0075889271683990955, + 0.01691439002752304, + 0.015024857595562935, + 0.008225045166909695, + 0.0173866655677557, + 0.00936823058873415, + 0.008413758128881454, + 0.013542957603931427, + 0.007311159744858742, + 0.0074232653714716434, + 0.010079770348966122, + 0.015901973471045494, + 0.015018780715763569, + 0.008138768374919891, + 0.0062480829656124115, + 0.022898053750395775, + 0.0059865317307412624, + 0.004818007815629244, + 0.00946091115474701, + 0.005805289372801781, + 0.027645956724882126, + 0.00754575477913022, + 0.012377717532217503, + 0.010606970638036728, + 0.053228896111249924, + 0.007877511903643608, + 0.010023186914622784, + 0.01520529668778181, + 0.03246975690126419, + 0.00750306062400341, + 0.003340262919664383, + 0.003886659862473607, + 0.008395427837967873, + 0.003549502696841955, + 0.0046820188872516155, + 0.005419897846877575, + 0.0076640816405415535, + 0.00319363409653306, + 0.010367296636104584, + 0.013299593701958656, + 0.013152515515685081, + 0.010727990418672562, + 0.007130481768399477, + 0.0121225630864501, + 0.008759042248129845, + 0.00585513561964035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.026818417012691498 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014217512682080269 + ], + "opposite_logprob_tokens": [ + -4.264217376708984 + ], + "logprob_reward": -0.014217512682080269, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014217512682080269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32224297523498535, + 0.42033952474594116, + 0.032048825174570084, + 0.08960995823144913, + 0.03544082120060921, + 0.011462012305855751, + 0.04122697189450264, + 0.01566319353878498, + 0.01072071772068739, + 0.021245019510388374 + ], + "credit_map": [ + 0.02626875229179859, + 0.03334644436836243, + 0.018161507323384285, + 0.04093025624752045, + 0.07341702282428741, + 0.041896477341651917, + 0.053421031683683395, + 0.03480147570371628, + 0.037015993148088455, + 0.04092814028263092, + 0.03213866055011749, + 0.039830636233091354, + 0.0733686089515686, + 0.05683840066194534, + 0.06083820387721062, + 0.07938090711832047, + 0.0035928196739405394, + 0.0033270008862018585, + 0.003739967942237854, + 0.002971646375954151, + 0.008233136497437954, + 0.002742831828072667, + 0.004290231037884951, + 0.003151191398501396, + 0.0344841293990612, + 0.007838387973606586, + 0.010587546974420547, + 0.005016115494072437, + 0.006318176165223122, + 0.007245290093123913, + 0.010789796710014343, + 0.0073305172845721245, + 0.0026573557406663895, + 0.004055303987115622, + 0.0031224070116877556, + 0.005684297066181898, + 0.004381934180855751, + 0.004201921168714762, + 0.006098357029259205, + 0.005239245947450399, + 0.0012821214040741324, + 0.001364680822007358, + 0.0012257341295480728, + 0.0012029822682961822, + 0.0014619356952607632, + 0.002339835511520505, + 0.001479391474276781, + 0.001105330535210669, + 0.003415357554331422, + 0.009986767545342445, + 0.0027437927201390266, + 0.002409858163446188, + 0.004280945751816034, + 0.0025298132095485926, + 0.012371206656098366, + 0.00348923122510314, + 0.0014548730105161667, + 0.001966083189472556, + 0.0022644400596618652, + 0.0015376325463876128, + 0.0022751707583665848, + 0.0030508290510624647, + 0.0017633856041356921, + 0.0013507793191820383, + 0.0007773203542456031, + 0.000570664880797267, + 0.002129786415025592, + 0.0008421230595558882, + 0.0008320038323290646, + 0.0016974564641714096, + 0.002657693810760975, + 0.0012136694276705384, + 0.0026358491741120815, + 0.002216635039076209, + 0.002762583550065756, + 0.002458662260323763, + 0.0017653437098488212, + 0.0048067751340568066, + 0.0026536460500210524, + 0.0019455240108072758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.014217512682080269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033704668283462524 + ], + "opposite_logprob_tokens": [ + -3.4087047576904297 + ], + "logprob_reward": -0.033704668283462524, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033704668283462524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19909536838531494, + 0.12337949872016907, + 0.0591040700674057, + 0.11490202695131302, + 0.04489746317267418, + 0.054138824343681335, + 0.10967657715082169, + 0.11021140217781067, + 0.08738204836845398, + 0.09721274673938751 + ], + "credit_map": [ + 0.017292143777012825, + 0.022331498563289642, + 0.024738147854804993, + 0.04533439129590988, + 0.02376202680170536, + 0.022039664909243584, + 0.02645557001233101, + 0.017141925171017647, + 0.014767229557037354, + 0.010140478610992432, + 0.014834140427410603, + 0.020639969035983086, + 0.019968608394265175, + 0.009516103193163872, + 0.015803392976522446, + 0.017709583044052124, + 0.0049327644519507885, + 0.00654999166727066, + 0.004637249279767275, + 0.005112475249916315, + 0.01812037080526352, + 0.006220875773578882, + 0.00881278421729803, + 0.004717554897069931, + 0.02501634694635868, + 0.013378052972257137, + 0.007678793743252754, + 0.005476820282638073, + 0.02373795211315155, + 0.016395321115851402, + 0.014579443261027336, + 0.008639298379421234, + 0.0038434872403740883, + 0.006408453453332186, + 0.003542036982253194, + 0.009267428889870644, + 0.005720463581383228, + 0.00464347330853343, + 0.007105100899934769, + 0.004367016721516848, + 0.0056998576037585735, + 0.009557201527059078, + 0.004710769280791283, + 0.00487244687974453, + 0.004985125735402107, + 0.008457967080175877, + 0.010785752907395363, + 0.005069703329354525, + 0.007209050469100475, + 0.02965034358203411, + 0.006489004008471966, + 0.0054074800573289394, + 0.010660874657332897, + 0.008363495580852032, + 0.03356047719717026, + 0.008335846476256847, + 0.009934600442647934, + 0.007151767611503601, + 0.04128563776612282, + 0.005744334775954485, + 0.00793723575770855, + 0.008030424825847149, + 0.02504609525203705, + 0.00508130481466651, + 0.005704804789274931, + 0.007669783663004637, + 0.01631341129541397, + 0.006943997927010059, + 0.007569612003862858, + 0.013330451212823391, + 0.0199230145663023, + 0.009926971979439259, + 0.013048207387328148, + 0.015466897748410702, + 0.0124262236058712, + 0.01549777016043663, + 0.007946747355163097, + 0.015035190619528294, + 0.010416100732982159, + 0.007375601679086685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.033704668283462524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059483409859240055 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.0059483409859240055, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059483409859240055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11242709308862686, + 0.31032517552375793, + 0.035045307129621506, + 0.08614278584718704, + 0.053697340190410614, + 0.05532750114798546, + 0.08194586634635925, + 0.06940935552120209, + 0.06621836125850677, + 0.1294611394405365 + ], + "credit_map": [ + 0.009894176386296749, + 0.013376172631978989, + 0.008543347008526325, + 0.018057914450764656, + 0.02223961614072323, + 0.012813945300877094, + 0.015811584889888763, + 0.011690338142216206, + 0.028066275641322136, + 0.031883690506219864, + 0.026552600786089897, + 0.03474685549736023, + 0.04656459391117096, + 0.04134499281644821, + 0.04318872094154358, + 0.057977449148893356, + 0.0037316891830414534, + 0.0041203550063073635, + 0.003263462334871292, + 0.0027194798458367586, + 0.009374150075018406, + 0.0034242430701851845, + 0.004719938151538372, + 0.003691987367346883, + 0.027912046760320663, + 0.008142593316733837, + 0.00825436133891344, + 0.004720850382000208, + 0.011588274501264095, + 0.006892335135489702, + 0.011054880917072296, + 0.007577436976134777, + 0.008142496459186077, + 0.005897049326449633, + 0.00379259348846972, + 0.006759351585060358, + 0.006558561231940985, + 0.005929578095674515, + 0.010461153462529182, + 0.006156555376946926, + 0.004862763918936253, + 0.010642234236001968, + 0.004334505647420883, + 0.006349741015583277, + 0.00611034594476223, + 0.008415032178163528, + 0.00981749128550291, + 0.004795389715582132, + 0.005365637596696615, + 0.021785231307148933, + 0.004143387544900179, + 0.0049042063765227795, + 0.009864314459264278, + 0.005009972024708986, + 0.024001212790608406, + 0.006871897727251053, + 0.004347473848611116, + 0.006210976280272007, + 0.021838778629899025, + 0.005438316147774458, + 0.0062205432914197445, + 0.006281757261604071, + 0.014806846156716347, + 0.004264666233211756, + 0.005625056568533182, + 0.003711097640916705, + 0.012893368490040302, + 0.004439996555447578, + 0.007211684249341488, + 0.011908940970897675, + 0.012336711399257183, + 0.00809150654822588, + 0.01976296864449978, + 0.014546472579240799, + 0.01683260127902031, + 0.017377451062202454, + 0.012071377597749233, + 0.02160082384943962, + 0.015113704837858677, + 0.012155728414654732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0059483409859240055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020572498440742493 + ], + "opposite_logprob_tokens": [ + -3.8955724239349365 + ], + "logprob_reward": -0.020572498440742493, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020572498440742493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15203602612018585, + 0.2021227926015854, + 0.12487698346376419, + 0.05690743401646614, + 0.09538274258375168, + 0.07022146135568619, + 0.045892391353845596, + 0.08993002772331238, + 0.04474850371479988, + 0.11788168549537659 + ], + "credit_map": [ + 0.012358216568827629, + 0.014751171693205833, + 0.01106620579957962, + 0.01687304489314556, + 0.033552736043930054, + 0.02320457249879837, + 0.024102818220853806, + 0.01612725295126438, + 0.01994769647717476, + 0.019237030297517776, + 0.018140237778425217, + 0.02703551948070526, + 0.0393730066716671, + 0.022106798365712166, + 0.027016961947083473, + 0.02926553785800934, + 0.011626661755144596, + 0.014802570454776287, + 0.011557864025235176, + 0.013336420990526676, + 0.03331538289785385, + 0.011272952891886234, + 0.018241476267576218, + 0.010723643936216831, + 0.01830676943063736, + 0.005554467439651489, + 0.006689868867397308, + 0.003981974441558123, + 0.007203044835478067, + 0.0040723709389567375, + 0.006049886345863342, + 0.005049051716923714, + 0.012785500846803188, + 0.009187515825033188, + 0.007322665303945541, + 0.016662850975990295, + 0.008685902692377567, + 0.01014429796487093, + 0.02285701222717762, + 0.0077369846403598785, + 0.009813496842980385, + 0.01088106818497181, + 0.007851722650229931, + 0.006775823887437582, + 0.009003764018416405, + 0.006036024075001478, + 0.01103017758578062, + 0.008829377591609955, + 0.006272405385971069, + 0.005519020836800337, + 0.005290957633405924, + 0.004965587053447962, + 0.004629882518202066, + 0.005697356071323156, + 0.008705521933734417, + 0.004811657126992941, + 0.006100856699049473, + 0.01033747661858797, + 0.018274718895554543, + 0.00769664766266942, + 0.009386040270328522, + 0.013949225656688213, + 0.016306031495332718, + 0.007879018783569336, + 0.003342946758493781, + 0.0029557247180491686, + 0.008391917683184147, + 0.0028803686145693064, + 0.00433914503082633, + 0.007836705073714256, + 0.01026090420782566, + 0.0047407932579517365, + 0.015259752050042152, + 0.009814384393393993, + 0.009098711423575878, + 0.01328864973038435, + 0.009422164410352707, + 0.03477795422077179, + 0.012081820517778397, + 0.014138240367174149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.020572498440742493 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dynamic shot reminiscent of Spider-Man: Into the Spider-Verse, the camera zooms forward at breakneck speed, capturing an absurd close-up. The subject is a cartoon character with a distinctly clay-like texture, its surface bearing subtle finger impressions, imbued with a handmade quality. Clad in a vibrant, saturated yellow jumpsuit, the figure tilts its head back, eyes wide open, an expression of innocent, focused anticipation. It is set against a backdrop woven from surreal pop art and a constantly flickering glitch aesthetic. Large, flatly painted geometric shapes interweave, rip, and reassemble with comic-book-like Ben-Day dots, filling the space with shifted chromatic aberrations and pixelated digital interference. The character holds aloft a cylindrical can of potato chips, the label clearly inscribed with \"Yogurt Flavor\" in bold green font. He decisively inverts the can, directly above his head. With one steady movement, the can's entire contents, complete with saddle-shaped chips, cascade down like an eerie waterfall, accompanied by a fine burst of green seasoning powder, instantly covering his entire head and face. The rapidly zoomed-in camera captures the shattering of each chip and the cascading of dust, blending seamlessly with the background's visual noise, ultimately framing the character's comical face after being \"cleaned\" by the chips, creating a chaotic and visually striking atmosphere of psychedelic comedy.", + "seed": 22278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_10/index_6748_seed_22278_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": -0.01546135190874338, + "vlm_logprob": -0.01546135190874338, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.01546135190874338, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01869124174118042 + ], + "opposite_logprob_tokens": [ + -4.518691062927246 + ], + "logprob_reward": -0.01869124174118042, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01869124174118042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0548701174557209, + 0.056079473346471786, + 0.035854846239089966, + 0.04534843936562538, + 0.46262919902801514, + 0.10988867282867432, + 0.043686822056770325, + 0.052187345921993256, + 0.06569644063711166, + 0.0737585723400116 + ], + "credit_map": [ + 0.003956540487706661, + 0.0033710645511746407, + 0.0063298605382442474, + 0.012894028797745705, + 0.008856218308210373, + 0.006270490121096373, + 0.010350313037633896, + 0.002841606270521879, + 0.009412319399416447, + 0.006921815685927868, + 0.004217551089823246, + 0.008229163475334644, + 0.009278948418796062, + 0.005729458294808865, + 0.0072806417010724545, + 0.005009583663195372, + 0.005767379887402058, + 0.006066774018108845, + 0.0025498673785477877, + 0.00334731163457036, + 0.006660297513008118, + 0.002472814405336976, + 0.006769643165171146, + 0.0022207582369446754, + 0.003907901234924793, + 0.007421011105179787, + 0.0034267071168869734, + 0.007523561827838421, + 0.010647682473063469, + 0.002683657221496105, + 0.007094195578247309, + 0.00264372443780303, + 0.020817311480641365, + 0.017844129353761673, + 0.20040127635002136, + 0.02420046180486679, + 0.026688190177083015, + 0.012414846569299698, + 0.147551029920578, + 0.012712011113762856, + 0.013920406810939312, + 0.013656219467520714, + 0.011424431577324867, + 0.012041204608976841, + 0.027438657358288765, + 0.007192810066044331, + 0.01838872954249382, + 0.005826220382004976, + 0.005027300212532282, + 0.0049446518532931805, + 0.0036781139206141233, + 0.009034229442477226, + 0.006766320206224918, + 0.003152359277009964, + 0.006316971965134144, + 0.004766879137605429, + 0.004936456214636564, + 0.0044522411189973354, + 0.005463744513690472, + 0.009014774113893509, + 0.013947741128504276, + 0.003693487262353301, + 0.0075776162557303905, + 0.0031012920662760735, + 0.005550216883420944, + 0.0064546456560492516, + 0.00920960120856762, + 0.012921938672661781, + 0.011236212216317654, + 0.004300223663449287, + 0.011048717424273491, + 0.00497488584369421, + 0.006582299247384071, + 0.006773831322789192, + 0.008069535717368126, + 0.011753544211387634, + 0.010814656503498554, + 0.004991316236555576, + 0.017553716897964478, + 0.0072196777909994125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.01869124174118042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865045003592968 + ], + "opposite_logprob_tokens": [ + -4.758650302886963 + ], + "logprob_reward": -0.00865045003592968, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865045003592968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11107784509658813, + 0.060233619064092636, + 0.01671864464879036, + 0.025819825008511543, + 0.14314761757850647, + 0.29034844040870667, + 0.08845736086368561, + 0.1737680435180664, + 0.06592466682195663, + 0.024503929540514946 + ], + "credit_map": [ + 0.011807224713265896, + 0.015177537687122822, + 0.01353406347334385, + 0.02174437791109085, + 0.01417723298072815, + 0.0136945191770792, + 0.014669009484350681, + 0.0062738754786551, + 0.008807326667010784, + 0.007226241752505302, + 0.0051120310090482235, + 0.011777580715715885, + 0.010215782560408115, + 0.005121496506035328, + 0.008698162622749805, + 0.00327499583363533, + 0.0017093728529289365, + 0.002162078395485878, + 0.0018361058318987489, + 0.0022670403122901917, + 0.0034065113868564367, + 0.0010517697082832456, + 0.0024533108808100224, + 0.0018324555130675435, + 0.0024854387156665325, + 0.0031687153968960047, + 0.00267124455422163, + 0.006550843361765146, + 0.003227825742214918, + 0.0018576848087832332, + 0.004106184933334589, + 0.0017518886597827077, + 0.006776946131139994, + 0.005487116053700447, + 0.061388228088617325, + 0.008745200000703335, + 0.007566205691546202, + 0.004135362338274717, + 0.043788179755210876, + 0.005260383244603872, + 0.04738844931125641, + 0.0315144807100296, + 0.02406448684632778, + 0.029441270977258682, + 0.06646611541509628, + 0.014935886487364769, + 0.06285562366247177, + 0.013682112097740173, + 0.018977729603648186, + 0.007041531149297953, + 0.007261407561600208, + 0.01851007342338562, + 0.010790733620524406, + 0.006707229185849428, + 0.013100892305374146, + 0.0060677616856992245, + 0.019001364707946777, + 0.018308443948626518, + 0.024362150579690933, + 0.03231879323720932, + 0.02819543518126011, + 0.012634800747036934, + 0.02853855863213539, + 0.01040849182754755, + 0.005573471542447805, + 0.007494526915252209, + 0.008440319448709488, + 0.013131988234817982, + 0.009737844578921795, + 0.004120526369661093, + 0.0129531379789114, + 0.0044728550128638744, + 0.002999131102114916, + 0.0025797777343541384, + 0.002424652222543955, + 0.003340063150972128, + 0.00392943387851119, + 0.002437596209347248, + 0.0047238836996257305, + 0.0020693917758762836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.00865045003592968 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04036175459623337, + 0.03335920721292496, + 0.02095617912709713, + 0.03639042377471924, + 0.6698617935180664, + 0.028901435434818268, + 0.03461752459406853, + 0.018531793728470802, + 0.08686653524637222, + 0.030153341591358185 + ], + "credit_map": [ + 0.004511516075581312, + 0.004653471522033215, + 0.004685630090534687, + 0.005771172698587179, + 0.005336093250662088, + 0.0031222989782691, + 0.007080378942191601, + 0.005201193504035473, + 0.0069892494939267635, + 0.004118580836802721, + 0.0033752380404621363, + 0.0041290996596217155, + 0.004711765330284834, + 0.0024756896309554577, + 0.004498310852795839, + 0.0030612738337367773, + 0.0039456686936318874, + 0.0017469640588387847, + 0.0020983375143259764, + 0.0032085287384688854, + 0.0031372015364468098, + 0.001531633548438549, + 0.00377495470456779, + 0.001512889633886516, + 0.0030318244826048613, + 0.00662463903427124, + 0.003678496927022934, + 0.006306611467152834, + 0.0052087572403252125, + 0.003171720542013645, + 0.00553972739726305, + 0.0028286457527428865, + 0.030951080843806267, + 0.0309585053473711, + 0.27808770537376404, + 0.039461396634578705, + 0.04708371311426163, + 0.02034832164645195, + 0.20459911227226257, + 0.01837199181318283, + 0.0031913279090076685, + 0.004004789516329765, + 0.004141435492783785, + 0.004275912418961525, + 0.004551278427243233, + 0.0020824011880904436, + 0.004716683179140091, + 0.0019376070704311132, + 0.004512378014624119, + 0.0027781976386904716, + 0.0034845892805606127, + 0.006811788771301508, + 0.005186735186725855, + 0.002653059083968401, + 0.006092860829085112, + 0.003097914857789874, + 0.0014545177109539509, + 0.001483793486841023, + 0.002262078458443284, + 0.002428630832582712, + 0.003983707167208195, + 0.001405263552442193, + 0.004259014036506414, + 0.001254789181984961, + 0.0057332757860422134, + 0.008610150776803493, + 0.011570828966796398, + 0.014614593237638474, + 0.01621529832482338, + 0.004755852743983269, + 0.01958300732076168, + 0.005783530883491039, + 0.0023920664098113775, + 0.0021485427860170603, + 0.003532959846779704, + 0.004585130140185356, + 0.003996339160948992, + 0.0022423022892326117, + 0.008196246810257435, + 0.0030597541481256485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.01607857272028923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098587669432163 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098587669432163, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098587669432163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08459945023059845, + 0.07958775758743286, + 0.04376378282904625, + 0.0601932518184185, + 0.31322258710861206, + 0.05399773269891739, + 0.06622766703367233, + 0.04345157742500305, + 0.12424108386039734, + 0.13071511685848236 + ], + "credit_map": [ + 0.006660303566604853, + 0.007624207995831966, + 0.011050067842006683, + 0.015040237456560135, + 0.01333056390285492, + 0.009008634835481644, + 0.013034351170063019, + 0.00885109044611454, + 0.011888035573065281, + 0.010106890462338924, + 0.005036971997469664, + 0.013691605068743229, + 0.012206300161778927, + 0.009078104048967361, + 0.011450788006186485, + 0.006129065528512001, + 0.003781307488679886, + 0.0033866604790091515, + 0.0051022060215473175, + 0.006559749133884907, + 0.011989152058959007, + 0.002124956576153636, + 0.008217932656407356, + 0.0026018242351710796, + 0.004510596860200167, + 0.00642892811447382, + 0.006520077586174011, + 0.010777084156870842, + 0.010223900899291039, + 0.006745783146470785, + 0.010968312621116638, + 0.004018567968159914, + 0.01515506487339735, + 0.018022220581769943, + 0.1161554604768753, + 0.023731932044029236, + 0.02483951859176159, + 0.01186249777674675, + 0.09012294560670853, + 0.013332955539226532, + 0.008955270051956177, + 0.006169158965349197, + 0.005742815788835287, + 0.007860906422138214, + 0.00815152283757925, + 0.003324962919577956, + 0.010219709947705269, + 0.0035733934491872787, + 0.008780768141150475, + 0.004748871549963951, + 0.0059825461357831955, + 0.01775469072163105, + 0.009359090588986874, + 0.004287429619580507, + 0.010713700205087662, + 0.004600577987730503, + 0.004644354805350304, + 0.00420820526778698, + 0.004270612262189388, + 0.006757823750376701, + 0.007954447530210018, + 0.0029119697865098715, + 0.00962711963802576, + 0.0030770476441830397, + 0.008086993359029293, + 0.013909983448684216, + 0.015652887523174286, + 0.01925840601325035, + 0.02556471899151802, + 0.008329177275300026, + 0.025213954970240593, + 0.008224977180361748, + 0.009712927043437958, + 0.010834218934178352, + 0.013246681541204453, + 0.024996094405651093, + 0.017238222062587738, + 0.007380401715636253, + 0.03303603455424309, + 0.0142705412581563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.004098587669432163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a cartoon character with a clay-like texture featuring subtle finger impressions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a vibrant, saturated yellow jumpsuit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character hold a cylindrical can of potato chips labeled 'Yogurt Flavor' in bold green font?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character invert the can above their head, causing chips and green seasoning powder to cascade down and cover their face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of surreal pop art with flickering glitch aesthetics, geometric shapes, and Ben-Day dots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787907376885414 + ], + "opposite_logprob_tokens": [ + -3.529788017272949 + ], + "logprob_reward": -0.029787907376885414, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787907376885414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20213741064071655, + 0.03899694234132767, + 0.0433032363653183, + 0.15497533977031708, + 0.3264997899532318, + 0.032829172909259796, + 0.037701286375522614, + 0.0361313670873642, + 0.08038712292909622, + 0.047038350254297256 + ], + "credit_map": [ + 0.01939218118786812, + 0.02445986308157444, + 0.026875978335738182, + 0.03421696275472641, + 0.032700713723897934, + 0.02479688823223114, + 0.026967674493789673, + 0.012727141380310059, + 0.006128211040049791, + 0.0060617271810770035, + 0.0023651334922760725, + 0.003968511242419481, + 0.005576408933848143, + 0.004999017808586359, + 0.004916358273476362, + 0.004981572274118662, + 0.005826757755130529, + 0.0038666354957967997, + 0.0041779810562729836, + 0.004984494764357805, + 0.009001488797366619, + 0.004131067078560591, + 0.0067682149820029736, + 0.00454659154638648, + 0.013259999454021454, + 0.014838773757219315, + 0.02056461200118065, + 0.03813999518752098, + 0.020780108869075775, + 0.01108433399349451, + 0.025434255599975586, + 0.010873248800635338, + 0.015954021364450455, + 0.02724270336329937, + 0.11811099946498871, + 0.026097070425748825, + 0.030260903760790825, + 0.010186680592596531, + 0.08903227001428604, + 0.009615137241780758, + 0.003383138682693243, + 0.0036328649148344994, + 0.004913883283734322, + 0.004786397330462933, + 0.005588357802480459, + 0.0028821702580899, + 0.005252163391560316, + 0.002390193287283182, + 0.004426924046128988, + 0.004282791167497635, + 0.003088080557063222, + 0.006177137605845928, + 0.00721670500934124, + 0.003605466103181243, + 0.006396218668669462, + 0.0025079590268433094, + 0.0037284225691109896, + 0.004297913517802954, + 0.00423487089574337, + 0.00519575783982873, + 0.007254370488226414, + 0.0028583414386957884, + 0.005808125250041485, + 0.00275356019847095, + 0.008294976316392422, + 0.006157282739877701, + 0.010643502697348595, + 0.015166448429226875, + 0.012361285276710987, + 0.0056046596728265285, + 0.01596289873123169, + 0.006196067668497562, + 0.005533890798687935, + 0.003837297670543194, + 0.005999875254929066, + 0.007422694470733404, + 0.004127237945795059, + 0.0073158638551831245, + 0.006775543559342623, + 0.006025944370776415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.029787907376885414 + } + ], + "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/0000110/group_11/index_tensor([8549, 8549])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_11/index_tensor([8549, 8549])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa30684ed860277cb8b80503bb9c7b2fb2f0ae5e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_11/index_tensor([8549, 8549])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot shot from a slightly overhead angle focuses the viewer's attention on a softly overhead-lit scene. The focal point of the image is a hand clad in a snug-fitting black lambskin glove, its surface gleaming with an understated yet elegant sheen, the leather's grain clearly visible in the light. The fingers of this hand are slightly bent, with only the index finger extended. With an extremely slow, appreciative rhythm, its fingertips carefully trace the surface of a brand-new tire. The tire, a purest matte black, is covered in an intricate and precise geometric tread pattern, each groove edge razor-sharp as if it were new. As the gloved fingertip glides across, the soft leather deforms minutely at the point of contact, creating a striking textural contrast with the tough rubber. The light precisely outlines the three-dimensional contours of the tire blocks and the subtle drainage grooves, casting soft shadows in the deep ravines, and blurring the background into an unrecognizable dark color, thus pushing this tactile and focused interaction between hand and tire into absolute visual focus, creating a tranquil atmosphere that is almost obsessive about the beauty of craftsmanship, performance and materials.", + "seed": 21812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_11/index_8549_seed_21812_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": -0.14442725591361522, + "vlm_logprob": -0.14442725591361522, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.14442725591361522, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04785814881324768 + ], + "opposite_logprob_tokens": [ + -3.422858238220215 + ], + "logprob_reward": -0.04785814881324768, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04785814881324768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29729390144348145, + 0.09128518402576447, + 0.23449239134788513, + 0.059940509498119354, + 0.055362410843372345, + 0.05389004200696945, + 0.11936380714178085, + 0.014962311834096909, + 0.039009809494018555, + 0.0343996025621891 + ], + "credit_map": [ + 0.048392798751592636, + 0.04492984712123871, + 0.03199448063969612, + 0.0415223091840744, + 0.0210100207477808, + 0.06560399383306503, + 0.021827230229973793, + 0.022013239562511444, + 0.013398964889347553, + 0.010855214670300484, + 0.008345614187419415, + 0.014672710560262203, + 0.009075182490050793, + 0.015475702472031116, + 0.008192532695829868, + 0.011269262060523033, + 0.03612540662288666, + 0.022470228374004364, + 0.012816844508051872, + 0.01822967827320099, + 0.019973374903202057, + 0.03694067522883415, + 0.05505947396159172, + 0.03287670388817787, + 0.0075791035778820515, + 0.006546939257532358, + 0.003241718979552388, + 0.00496537983417511, + 0.010309068486094475, + 0.00795905850827694, + 0.010800807736814022, + 0.008538436144590378, + 0.008838629350066185, + 0.003444022499024868, + 0.004346869885921478, + 0.00856253132224083, + 0.0063589573837816715, + 0.006295423023402691, + 0.007153256796300411, + 0.0103627173230052, + 0.006874529179185629, + 0.008689802139997482, + 0.003141809022054076, + 0.006261703558266163, + 0.007870231755077839, + 0.005829479545354843, + 0.009542555548250675, + 0.005679933819919825, + 0.016173865646123886, + 0.012087654322385788, + 0.0067371949553489685, + 0.01445451658219099, + 0.016618944704532623, + 0.013902567327022552, + 0.01261052768677473, + 0.02677854150533676, + 0.0025196848437190056, + 0.0018522925674915314, + 0.0007489510462619364, + 0.0011980386916548014, + 0.0017206919146701694, + 0.0021686647087335587, + 0.0016912608407437801, + 0.0030627271626144648, + 0.006083874497562647, + 0.008118207566440105, + 0.0029015035834163427, + 0.0038424916565418243, + 0.004582694265991449, + 0.005631378851830959, + 0.003242387203499675, + 0.004607275128364563, + 0.004631765652447939, + 0.005071217194199562, + 0.003975769504904747, + 0.0034155964385718107, + 0.004618487320840359, + 0.004199282731860876, + 0.004322915803641081, + 0.004164570011198521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.04785814881324768 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18004439771175385 + ], + "opposite_logprob_tokens": [ + -1.805044412612915 + ], + "logprob_reward": -0.18004439771175385, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18004439771175385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1377161592245102, + 0.2279149889945984, + 0.1438370943069458, + 0.21550437808036804, + 0.07777398079633713, + 0.0523453950881958, + 0.06531946361064911, + 0.0330122634768486, + 0.024183813482522964, + 0.022392457351088524 + ], + "credit_map": [ + 0.025203747674822807, + 0.017768925055861473, + 0.009577726945281029, + 0.016712823882699013, + 0.015334703959524632, + 0.01623731106519699, + 0.012237677350640297, + 0.024643249809741974, + 0.03561937063932419, + 0.03460729867219925, + 0.013593406416475773, + 0.02356024831533432, + 0.025329148396849632, + 0.05441542714834213, + 0.0183237474411726, + 0.022466333582997322, + 0.0239066481590271, + 0.013641349039971828, + 0.008480682037770748, + 0.012525602243840694, + 0.011529596522450447, + 0.023481886833906174, + 0.03225623071193695, + 0.01801510527729988, + 0.025475887581706047, + 0.022374227643013, + 0.00997652392834425, + 0.016615619882941246, + 0.049287348985672, + 0.02917248010635376, + 0.03840422257781029, + 0.024198060855269432, + 0.00999078806489706, + 0.003744564251974225, + 0.005048045888543129, + 0.010389289818704128, + 0.014524439349770546, + 0.012270832434296608, + 0.011037054471671581, + 0.01076896209269762, + 0.007638611365109682, + 0.008500175550580025, + 0.005146501585841179, + 0.006702341604977846, + 0.0052324747666716576, + 0.004147697240114212, + 0.009553107433021069, + 0.005424481350928545, + 0.009669058956205845, + 0.00821772962808609, + 0.004332470241934061, + 0.00813325121998787, + 0.012844402343034744, + 0.006667341571301222, + 0.006598217878490686, + 0.008856991305947304, + 0.005064431112259626, + 0.0050058201886713505, + 0.0023365879897028208, + 0.004086510743945837, + 0.0037515079602599144, + 0.005266039632260799, + 0.002323163440451026, + 0.005178199149668217, + 0.0033589822705835104, + 0.0033172250259667635, + 0.0021374861244112253, + 0.003906366880983114, + 0.002902858192101121, + 0.0033268213737756014, + 0.0018131159013137221, + 0.003420959459617734, + 0.003804505802690983, + 0.003514977404847741, + 0.0020739687606692314, + 0.002683530794456601, + 0.002369262743741274, + 0.003317579161375761, + 0.0019995940383523703, + 0.0026290370151400566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.18004439771175385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034353286027908 + ], + "opposite_logprob_tokens": [ + -2.3503434658050537 + ], + "logprob_reward": -0.10034353286027908, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034353286027908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17126639187335968, + 0.1690487265586853, + 0.07676514983177185, + 0.12623706459999084, + 0.09967945516109467, + 0.11728325486183167, + 0.08178949356079102, + 0.05483497306704521, + 0.08018659800291061, + 0.022908871993422508 + ], + "credit_map": [ + 0.029067404568195343, + 0.026861604303121567, + 0.013401485979557037, + 0.018807582557201385, + 0.017610272392630577, + 0.025183171033859253, + 0.012793031521141529, + 0.027541829273104668, + 0.026440730318427086, + 0.02457445114850998, + 0.01071796752512455, + 0.01700998842716217, + 0.017581896856427193, + 0.0423879474401474, + 0.013318641111254692, + 0.017017114907503128, + 0.018889935687184334, + 0.006795292720198631, + 0.0037578768096864223, + 0.005733179859817028, + 0.0048771752044558525, + 0.008155412040650845, + 0.01652812398970127, + 0.012028152123093605, + 0.019015684723854065, + 0.012342112138867378, + 0.008137678727507591, + 0.010263212025165558, + 0.023352159187197685, + 0.01626715436577797, + 0.017841611057519913, + 0.01901746727526188, + 0.015806609764695168, + 0.005637046881020069, + 0.0070603592321276665, + 0.014195695519447327, + 0.014076669700443745, + 0.013839555904269218, + 0.013534797355532646, + 0.015528718940913677, + 0.015812166035175323, + 0.017654215916991234, + 0.008653082884848118, + 0.015999197959899902, + 0.01231929287314415, + 0.01356169767677784, + 0.02022482454776764, + 0.013058770447969437, + 0.012687234207987785, + 0.009700390510261059, + 0.005260447971522808, + 0.006528032012283802, + 0.01877782680094242, + 0.00934648234397173, + 0.00862154271453619, + 0.01086754072457552, + 0.00898808054625988, + 0.006017029285430908, + 0.005442660301923752, + 0.007447257172316313, + 0.006298171356320381, + 0.006602085195481777, + 0.006020248867571354, + 0.008019444532692432, + 0.01227165199816227, + 0.014277462847530842, + 0.004827359691262245, + 0.00939622800797224, + 0.009499894455075264, + 0.0100663797929883, + 0.006646371446549892, + 0.013201247900724411, + 0.0035534179769456387, + 0.003139645094051957, + 0.002008734969422221, + 0.0027014813385903835, + 0.003008873201906681, + 0.0033705926034599543, + 0.002494659973308444, + 0.0026314686983823776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.10034353286027908 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38714489340782166 + ], + "opposite_logprob_tokens": [ + -1.137144923210144 + ], + "logprob_reward": -0.38714489340782166, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38714489340782166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26190516352653503, + 0.0926096960902214, + 0.10718829184770584, + 0.06556829065084457, + 0.10197756439447403, + 0.1300145536661148, + 0.06325655430555344, + 0.05401775240898132, + 0.05583658069372177, + 0.06762556731700897 + ], + "credit_map": [ + 0.04136325418949127, + 0.034390661865472794, + 0.020575718954205513, + 0.0381009504199028, + 0.0256723091006279, + 0.046584539115428925, + 0.019885776564478874, + 0.03533197566866875, + 0.015794742852449417, + 0.01029223669320345, + 0.006907869596034288, + 0.008973286487162113, + 0.009536475874483585, + 0.02100846916437149, + 0.006673875730484724, + 0.013422733172774315, + 0.015614161267876625, + 0.012845804914832115, + 0.006380598060786724, + 0.010525599122047424, + 0.013334055431187153, + 0.014630657620728016, + 0.018903030082583427, + 0.014954381622374058, + 0.00808397401124239, + 0.007896537892520428, + 0.003957353997975588, + 0.0079905791208148, + 0.010208205319941044, + 0.007817461155354977, + 0.011086549609899521, + 0.008527633734047413, + 0.017660729587078094, + 0.0042581986635923386, + 0.006572754122316837, + 0.01147178653627634, + 0.013769423589110374, + 0.01349011529237032, + 0.016525551676750183, + 0.01822899654507637, + 0.019079478457570076, + 0.020104840397834778, + 0.009473979473114014, + 0.017429091036319733, + 0.015570376999676228, + 0.015325877815485, + 0.02011100761592388, + 0.012919891625642776, + 0.009531469084322453, + 0.007780337706208229, + 0.005236275028437376, + 0.006681439001113176, + 0.01153773907572031, + 0.006634450983256102, + 0.00775481341406703, + 0.008100029081106186, + 0.007262750994414091, + 0.007696879096329212, + 0.005663393996655941, + 0.010063591413199902, + 0.00493208970874548, + 0.006326678209006786, + 0.0045356485061347485, + 0.007536720484495163, + 0.006966727785766125, + 0.009327064268290997, + 0.0052298903465271, + 0.009203018620610237, + 0.007976911030709743, + 0.006854139734059572, + 0.004799520131200552, + 0.005479307845234871, + 0.01010409090667963, + 0.011643124744296074, + 0.0064478120766580105, + 0.012945962138473988, + 0.0057251895777881145, + 0.01068002451211214, + 0.005803624168038368, + 0.004275738261640072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.38714489340782166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745306774973869 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745306774973869, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745306774973869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1702709048986435, + 0.12354996800422668, + 0.08864281326532364, + 0.13661110401153564, + 0.13306684792041779, + 0.05913330987095833, + 0.11020024865865707, + 0.1284453272819519, + 0.026346543803811073, + 0.023732885718345642 + ], + "credit_map": [ + 0.03396713361144066, + 0.017322881147265434, + 0.010353297926485538, + 0.01621248759329319, + 0.01932218298316002, + 0.03613593429327011, + 0.01205599494278431, + 0.024900998920202255, + 0.021643156185746193, + 0.01638256572186947, + 0.007355430629104376, + 0.015890201553702354, + 0.014072301797568798, + 0.024393176659941673, + 0.011052383109927177, + 0.01276074443012476, + 0.01500479131937027, + 0.00920924637466669, + 0.005524761974811554, + 0.008487775921821594, + 0.009335504844784737, + 0.010977976024150848, + 0.017324741929769516, + 0.012778017669916153, + 0.017086846753954887, + 0.011908181942999363, + 0.007658516056835651, + 0.010964169166982174, + 0.027958901599049568, + 0.01791083812713623, + 0.021357571706175804, + 0.021766090765595436, + 0.02125508524477482, + 0.009722625836730003, + 0.012664077803492546, + 0.01910427026450634, + 0.012108877301216125, + 0.014497335068881512, + 0.020127281546592712, + 0.0235872995108366, + 0.008362554013729095, + 0.010103104636073112, + 0.0038577569648623466, + 0.005035367328673601, + 0.007579485420137644, + 0.006470042280852795, + 0.010281775146722794, + 0.007443226873874664, + 0.015185066498816013, + 0.013382617384195328, + 0.0056397090665996075, + 0.008591385558247566, + 0.0241598729044199, + 0.015214014798402786, + 0.011044437997043133, + 0.016983140259981155, + 0.018703943118453026, + 0.021939989179372787, + 0.013574707321822643, + 0.02040822245180607, + 0.009324013255536556, + 0.012784121558070183, + 0.010819727554917336, + 0.020890597254037857, + 0.004365754313766956, + 0.004029002506285906, + 0.0026404368691146374, + 0.0037240053061395884, + 0.0022601322270929813, + 0.0038057316560298204, + 0.0021464000456035137, + 0.0033750790171325207, + 0.0037606563419103622, + 0.004201793577522039, + 0.0021929852664470673, + 0.0033679837360978127, + 0.002636194694787264, + 0.003376971697434783, + 0.0018560111057013273, + 0.0023402876686304808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.006745306774973869 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot shot from a slightly overhead angle focuses the viewer's attention on a softly overhead-lit scene. The focal point of the image is a hand clad in a snug-fitting black lambskin glove, its surface gleaming with an understated yet elegant sheen, the leather's grain clearly visible in the light. The fingers of this hand are slightly bent, with only the index finger extended. With an extremely slow, appreciative rhythm, its fingertips carefully trace the surface of a brand-new tire. The tire, a purest matte black, is covered in an intricate and precise geometric tread pattern, each groove edge razor-sharp as if it were new. As the gloved fingertip glides across, the soft leather deforms minutely at the point of contact, creating a striking textural contrast with the tough rubber. The light precisely outlines the three-dimensional contours of the tire blocks and the subtle drainage grooves, casting soft shadows in the deep ravines, and blurring the background into an unrecognizable dark color, thus pushing this tactile and focused interaction between hand and tire into absolute visual focus, creating a tranquil atmosphere that is almost obsessive about the beauty of craftsmanship, performance and materials.", + "seed": 21813, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_11/index_8549_seed_21813_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": -0.11460893684998155, + "vlm_logprob": -0.11460893684998155, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.11460893684998155, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040846217423677444 + ], + "opposite_logprob_tokens": [ + -3.540846109390259 + ], + "logprob_reward": -0.040846217423677444, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.040846217423677444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020185595378279686, + 0.017012596130371094, + 0.025873715057969093, + 0.04464248567819595, + 0.05313616618514061, + 0.06484681367874146, + 0.12400457262992859, + 0.2736019492149353, + 0.3200848400592804, + 0.05661123991012573 + ], + "credit_map": [ + 0.002953606890514493, + 0.0029976861551404, + 0.0032869598362594843, + 0.002722112461924553, + 0.001349576748907566, + 0.0014806969556957483, + 0.0020169364288449287, + 0.003378019668161869, + 0.0037534653674811125, + 0.0021261798683553934, + 0.0028122994117438793, + 0.0015283121028915048, + 0.0009334614733234048, + 0.001348935766145587, + 0.002242116257548332, + 0.0022678272798657417, + 0.0038766837678849697, + 0.0036714551970362663, + 0.0045720962807536125, + 0.0029215847607702017, + 0.0015993465203791857, + 0.002003479516133666, + 0.0033074140083044767, + 0.003921654541045427, + 0.010787020437419415, + 0.006259556859731674, + 0.006314254831522703, + 0.004596095532178879, + 0.0021616581361740828, + 0.0028409743681550026, + 0.004089871421456337, + 0.007593054324388504, + 0.007073097862303257, + 0.005946916528046131, + 0.008089397102594376, + 0.005854051560163498, + 0.0025104880332946777, + 0.004019627347588539, + 0.006569666787981987, + 0.013072923757135868, + 0.006838623434305191, + 0.004687402863055468, + 0.013083726167678833, + 0.006049891468137503, + 0.003666415810585022, + 0.00468415766954422, + 0.004722699988633394, + 0.021113893017172813, + 0.019066883251070976, + 0.0075996434316039085, + 0.02194160409271717, + 0.012250066734850407, + 0.0061464025638997555, + 0.006338770966976881, + 0.0100735267624259, + 0.04058767482638359, + 0.03012780472636223, + 0.029045673087239265, + 0.04163793474435806, + 0.04293634742498398, + 0.03277382627129555, + 0.01799020916223526, + 0.04274113103747368, + 0.03634902834892273, + 0.041828230023384094, + 0.07256325334310532, + 0.03423265367746353, + 0.04099813848733902, + 0.02585304155945778, + 0.021097905933856964, + 0.04010789096355438, + 0.043403711169958115, + 0.008006732910871506, + 0.011970479041337967, + 0.009846439585089684, + 0.007836727425456047, + 0.0037191163282841444, + 0.002926107030361891, + 0.00504093524068594, + 0.007264698389917612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.040846217423677444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14298459887504578 + ], + "opposite_logprob_tokens": [ + -2.017984628677368 + ], + "logprob_reward": -0.14298459887504578, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14298459887504578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03665915131568909, + 0.01709873043000698, + 0.015346845611929893, + 0.02909974381327629, + 0.05644267052412033, + 0.0829009935259819, + 0.031473223119974136, + 0.6076997518539429, + 0.07407869398593903, + 0.049200158566236496 + ], + "credit_map": [ + 0.0077008213847875595, + 0.005587655585259199, + 0.005554854869842529, + 0.004045821726322174, + 0.0015478505520150065, + 0.0021590900141745806, + 0.004603775683790445, + 0.005459281150251627, + 0.0030451565980911255, + 0.0018294764449819922, + 0.0030057448893785477, + 0.0014996322570368648, + 0.0016225987346842885, + 0.0015854078810662031, + 0.0021827942691743374, + 0.0023279194720089436, + 0.0035677598789334297, + 0.0021300148218870163, + 0.002874864963814616, + 0.0011524866567924619, + 0.0010165816638618708, + 0.0011040751123800874, + 0.0011238917941227555, + 0.0023771708365529776, + 0.004398877266794443, + 0.0032329948153346777, + 0.0046903942711651325, + 0.0020324401557445526, + 0.0011650248197838664, + 0.002056174911558628, + 0.0024097440764307976, + 0.009114094078540802, + 0.009124623611569405, + 0.007818318903446198, + 0.008228934369981289, + 0.0079640569165349, + 0.0029760869219899178, + 0.004053397104144096, + 0.004451698623597622, + 0.011825548484921455, + 0.011484840884804726, + 0.00816056877374649, + 0.015403023920953274, + 0.004900672473013401, + 0.0034918291494250298, + 0.00550581282004714, + 0.004384358413517475, + 0.02956988476216793, + 0.004827437456697226, + 0.0026461281813681126, + 0.005063818302005529, + 0.00532689830288291, + 0.00244358042255044, + 0.0015288017457351089, + 0.0025881652254611254, + 0.007048395462334156, + 0.09932965040206909, + 0.07189750671386719, + 0.11284980922937393, + 0.07793102413415909, + 0.1003984585404396, + 0.029667634516954422, + 0.05298407003283501, + 0.06264162808656693, + 0.0074478499591350555, + 0.018085693940520287, + 0.009813590906560421, + 0.009013732895255089, + 0.0063589103519916534, + 0.004531177692115307, + 0.008742648176848888, + 0.010085091926157475, + 0.0056246016174554825, + 0.005594272632151842, + 0.009084402583539486, + 0.0060672396793961525, + 0.006498167756944895, + 0.003306930884718895, + 0.0058280485682189465, + 0.00719649950042367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.14298459887504578 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016129195690155 + ], + "opposite_logprob_tokens": [ + -1.701612949371338 + ], + "logprob_reward": -0.2016129195690155, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016129195690155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03490472957491875, + 0.025898169726133347, + 0.01524145808070898, + 0.05086725205183029, + 0.07080452144145966, + 0.08565884828567505, + 0.17006179690361023, + 0.43770909309387207, + 0.07404670864343643, + 0.034807413816452026 + ], + "credit_map": [ + 0.006110761314630508, + 0.0046302881091833115, + 0.004807019140571356, + 0.004175575915724039, + 0.002847476629540324, + 0.0034931425470858812, + 0.0035476279444992542, + 0.005292840301990509, + 0.005586822051554918, + 0.0031838417053222656, + 0.004697086755186319, + 0.00201538042165339, + 0.0021896162070333958, + 0.0021393699571490288, + 0.0027368857990950346, + 0.0033491672948002815, + 0.0018983653280884027, + 0.0021926483605057, + 0.001990969991311431, + 0.0018852268112823367, + 0.0008389578433707356, + 0.0013437949819490314, + 0.0018157192971557379, + 0.0032757774461060762, + 0.009120662696659565, + 0.00503799831494689, + 0.007350569125264883, + 0.003053341992199421, + 0.002502569230273366, + 0.004314193036407232, + 0.0031913709826767445, + 0.016296550631523132, + 0.012405604124069214, + 0.008434910327196121, + 0.011028137058019638, + 0.011044404469430447, + 0.0033368603326380253, + 0.00471142865717411, + 0.0051935953088104725, + 0.014649583958089352, + 0.012362463399767876, + 0.007275228854268789, + 0.015485611744225025, + 0.007340218871831894, + 0.003317704424262047, + 0.005715383216738701, + 0.004743431694805622, + 0.029418816789984703, + 0.021971561014652252, + 0.011539488099515438, + 0.025716014206409454, + 0.019646866247057915, + 0.012331465259194374, + 0.010319724678993225, + 0.030532848089933395, + 0.03800384700298309, + 0.0437421016395092, + 0.0858977809548378, + 0.05592641979455948, + 0.1302022486925125, + 0.03194974362850189, + 0.017593106254935265, + 0.025781404227018356, + 0.046616289764642715, + 0.009958354756236076, + 0.017494795843958855, + 0.009810395538806915, + 0.009212164208292961, + 0.0060015516355633736, + 0.004248747136443853, + 0.008464112877845764, + 0.008856591768562794, + 0.004984749481081963, + 0.004450784996151924, + 0.003897290211170912, + 0.00442579435184598, + 0.004104330204427242, + 0.002573112491518259, + 0.004493444226682186, + 0.005877907387912273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.2016129195690155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179968923330307 + ], + "opposite_logprob_tokens": [ + -1.8049689531326294 + ], + "logprob_reward": -0.179968923330307, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179968923330307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11594467610120773, + 0.0464257076382637, + 0.0667845606803894, + 0.17791704833507538, + 0.03915945813059807, + 0.10237330198287964, + 0.05238451063632965, + 0.23610380291938782, + 0.09242282807826996, + 0.07048410922288895 + ], + "credit_map": [ + 0.019586725160479546, + 0.017190836369991302, + 0.017345914617180824, + 0.010092617012560368, + 0.007758020889014006, + 0.009038081392645836, + 0.013494464568793774, + 0.02143801935017109, + 0.006979143712669611, + 0.004082888830453157, + 0.006323230918496847, + 0.00572515744715929, + 0.005150642246007919, + 0.004846405703574419, + 0.005885621532797813, + 0.00743261631578207, + 0.011319348588585854, + 0.008106512948870659, + 0.011618190445005894, + 0.0073537337593734264, + 0.004870151169598103, + 0.00538606196641922, + 0.007090114522725344, + 0.011040449142456055, + 0.029788125306367874, + 0.018761316314339638, + 0.025432629510760307, + 0.012200282886624336, + 0.011013084091246128, + 0.016134988516569138, + 0.01336689107120037, + 0.05121972784399986, + 0.006961777340620756, + 0.005441701039671898, + 0.004711831919848919, + 0.005766174755990505, + 0.002030376810580492, + 0.0028510577976703644, + 0.00492467824369669, + 0.006471862550824881, + 0.01237332820892334, + 0.010707587003707886, + 0.01848294585943222, + 0.006522015668451786, + 0.0047685205936431885, + 0.007908347994089127, + 0.0061024390161037445, + 0.03550812602043152, + 0.007298337761312723, + 0.0059147491119802, + 0.007764520589262247, + 0.008479341864585876, + 0.003968344535678625, + 0.0028386113699525595, + 0.006872399244457483, + 0.009248205460608006, + 0.03386056050658226, + 0.02192656323313713, + 0.04061965271830559, + 0.04577406868338585, + 0.03840237483382225, + 0.010274591855704784, + 0.018013635650277138, + 0.02723236009478569, + 0.011972134001553059, + 0.02051180601119995, + 0.014574447646737099, + 0.01219131238758564, + 0.007596057839691639, + 0.0051123397424817085, + 0.009133316576480865, + 0.011331419460475445, + 0.008088231086730957, + 0.009568496607244015, + 0.008337804116308689, + 0.010487551800906658, + 0.008625212125480175, + 0.005856077186763287, + 0.0111399469897151, + 0.008380787447094917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.179968923330307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632025051862001 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632025051862001, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632025051862001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09658457338809967, + 0.059489358216524124, + 0.02734152041375637, + 0.025190532207489014, + 0.08751881867647171, + 0.12138481438159943, + 0.05194063112139702, + 0.2685597836971283, + 0.20738154649734497, + 0.054608382284641266 + ], + "credit_map": [ + 0.020426491275429726, + 0.013395139016211033, + 0.014234252274036407, + 0.006333943456411362, + 0.005248833913356066, + 0.006612563505768776, + 0.010233405977487564, + 0.02009994350373745, + 0.012393181212246418, + 0.008334857411682606, + 0.010212118737399578, + 0.005707832053303719, + 0.003003856400027871, + 0.004325724672526121, + 0.007300649769604206, + 0.008211139589548111, + 0.0039096400141716, + 0.003911236301064491, + 0.0051730601117014885, + 0.0029450322035700083, + 0.002093897433951497, + 0.002057325327768922, + 0.0024815129581838846, + 0.004769816994667053, + 0.002780207432806492, + 0.003720724955201149, + 0.004220221191644669, + 0.0028774046804755926, + 0.0015546752838417888, + 0.00185815105214715, + 0.0028939591720700264, + 0.005285189021378756, + 0.014435432851314545, + 0.013209977187216282, + 0.011664626188576221, + 0.011893532238900661, + 0.005147355142980814, + 0.0054692551493644714, + 0.010268316604197025, + 0.015430324710905552, + 0.01715797744691372, + 0.012094501405954361, + 0.020855994895100594, + 0.012509597465395927, + 0.006524308118969202, + 0.007825357839465141, + 0.008615612052381039, + 0.03580145537853241, + 0.004748365841805935, + 0.004796015564352274, + 0.007648138329386711, + 0.00637776916846633, + 0.004456048831343651, + 0.0035059740766882896, + 0.011345796287059784, + 0.009062527678906918, + 0.03093002177774906, + 0.026303844526410103, + 0.0433601550757885, + 0.0446317195892334, + 0.044411975890398026, + 0.015198319219052792, + 0.028515610843896866, + 0.03520815074443817, + 0.021670380607247353, + 0.052337996661663055, + 0.02756640873849392, + 0.024973874911665916, + 0.016599904745817184, + 0.011773334816098213, + 0.02373427338898182, + 0.028725385665893555, + 0.01068198587745428, + 0.006502896081656218, + 0.008148133754730225, + 0.007137231528759003, + 0.005379508249461651, + 0.0030467251781374216, + 0.007342007476836443, + 0.006369893439114094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.007632025051862001 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot shot from a slightly overhead angle focuses the viewer's attention on a softly overhead-lit scene. The focal point of the image is a hand clad in a snug-fitting black lambskin glove, its surface gleaming with an understated yet elegant sheen, the leather's grain clearly visible in the light. The fingers of this hand are slightly bent, with only the index finger extended. With an extremely slow, appreciative rhythm, its fingertips carefully trace the surface of a brand-new tire. The tire, a purest matte black, is covered in an intricate and precise geometric tread pattern, each groove edge razor-sharp as if it were new. As the gloved fingertip glides across, the soft leather deforms minutely at the point of contact, creating a striking textural contrast with the tough rubber. The light precisely outlines the three-dimensional contours of the tire blocks and the subtle drainage grooves, casting soft shadows in the deep ravines, and blurring the background into an unrecognizable dark color, thus pushing this tactile and focused interaction between hand and tire into absolute visual focus, creating a tranquil atmosphere that is almost obsessive about the beauty of craftsmanship, performance and materials.", + "seed": 21814, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_11/index_8549_seed_21814_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": -0.05645299460738897, + "vlm_logprob": -0.05645299460738897, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.05645299460738897, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03166840225458145 + ], + "opposite_logprob_tokens": [ + -4.531668186187744 + ], + "logprob_reward": -0.03166840225458145, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03166840225458145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0747910737991333, + 0.06949066370725632, + 0.08044105023145676, + 0.0629325732588768, + 0.058756306767463684, + 0.27190807461738586, + 0.23836538195610046, + 0.032862309366464615, + 0.04744891822338104, + 0.06300368160009384 + ], + "credit_map": [ + 0.011767400428652763, + 0.006081622093915939, + 0.0036586811766028404, + 0.006573152262717485, + 0.016199417412281036, + 0.010482154786586761, + 0.008489933796226978, + 0.011538704857230186, + 0.008705100044608116, + 0.0039442554116249084, + 0.003391862381249666, + 0.005653999280184507, + 0.006665348540991545, + 0.0193755105137825, + 0.009843803010880947, + 0.011910778470337391, + 0.010298016481101513, + 0.00851327646523714, + 0.005395474843680859, + 0.005969828460365534, + 0.0067258006893098354, + 0.01277218572795391, + 0.013368182815611362, + 0.01739828661084175, + 0.008819589391350746, + 0.0058432468213140965, + 0.004132019355893135, + 0.0049612452276051044, + 0.0034775903914123774, + 0.009811514057219028, + 0.008914765901863575, + 0.016972599551081657, + 0.007968785241246223, + 0.006713589187711477, + 0.004516048356890678, + 0.005361839197576046, + 0.00865352526307106, + 0.008057907223701477, + 0.00853239931166172, + 0.00895220972597599, + 0.04644900932908058, + 0.023473812267184258, + 0.016995558515191078, + 0.02065463177859783, + 0.03332136571407318, + 0.045137643814086914, + 0.03287181630730629, + 0.05300420522689819, + 0.02844640240073204, + 0.01645110920071602, + 0.009525351226329803, + 0.015530615113675594, + 0.052069347351789474, + 0.046344246715307236, + 0.038732513785362244, + 0.031265776604413986, + 0.0034239746164530516, + 0.002280367538332939, + 0.0016463943757116795, + 0.0020892114844173193, + 0.0025380828883498907, + 0.009057112038135529, + 0.003673459403216839, + 0.008153706789016724, + 0.005986875854432583, + 0.005718868691474199, + 0.003003437537699938, + 0.004308062139898539, + 0.002572060329839587, + 0.013360414654016495, + 0.004049894865602255, + 0.00844929926097393, + 0.006394537631422281, + 0.004854207392781973, + 0.004139969125390053, + 0.004887192510068417, + 0.013683388009667397, + 0.01011054776608944, + 0.009421749971807003, + 0.00951208733022213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.03166840225458145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500292778015137 + ], + "opposite_logprob_tokens": [ + -2.9300029277801514 + ], + "logprob_reward": -0.05500292778015137, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500292778015137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3960091173648834, + 0.06516524404287338, + 0.11183957010507584, + 0.05136779695749283, + 0.029529793187975883, + 0.10623908042907715, + 0.019137222319841385, + 0.06884130835533142, + 0.07842227816581726, + 0.07344857603311539 + ], + "credit_map": [ + 0.0487237311899662, + 0.02911512181162834, + 0.0215593334287405, + 0.02936801128089428, + 0.08794260025024414, + 0.06788011640310287, + 0.05441565811634064, + 0.05700456351041794, + 0.008499654941260815, + 0.005980793386697769, + 0.005215102341026068, + 0.004903819877654314, + 0.006868154741823673, + 0.015463549643754959, + 0.010023398324847221, + 0.008210772648453712, + 0.01831786520779133, + 0.0104849673807621, + 0.007264027372002602, + 0.009508711285889149, + 0.011978576891124249, + 0.019940944388508797, + 0.015340229496359825, + 0.019004248082637787, + 0.006897899322211742, + 0.006691372953355312, + 0.0036134340334683657, + 0.00470053032040596, + 0.004995069000869989, + 0.006852664519101381, + 0.00801231898367405, + 0.009604508057236671, + 0.004068542271852493, + 0.0035677137784659863, + 0.0025244811549782753, + 0.0033862253185361624, + 0.003980063367635012, + 0.003444315167143941, + 0.0038600312545895576, + 0.004698422271758318, + 0.018298732116818428, + 0.010340961627662182, + 0.007107546087354422, + 0.009167863987386227, + 0.012059072963893414, + 0.01553295087069273, + 0.013195830397307873, + 0.02053612656891346, + 0.002044045366346836, + 0.0014889163430780172, + 0.0012857933761551976, + 0.0014708727831020951, + 0.003016869304701686, + 0.004269344266504049, + 0.0027606647927314043, + 0.0028007167857140303, + 0.007692393846809864, + 0.004557158797979355, + 0.004280949477106333, + 0.004832737613469362, + 0.005208855960518122, + 0.014807473868131638, + 0.00855798926204443, + 0.018903745338320732, + 0.01947002485394478, + 0.005232873372733593, + 0.0035286264028400183, + 0.0034990226849913597, + 0.005542883183807135, + 0.024914881214499474, + 0.006369050592184067, + 0.009864919818937778, + 0.009603329934179783, + 0.005063439719378948, + 0.004469758365303278, + 0.0051777902990579605, + 0.009382608346641064, + 0.012785612605512142, + 0.010689833201467991, + 0.016276206821203232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.05500292778015137 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295662075281143 + ], + "opposite_logprob_tokens": [ + -2.2379565238952637 + ], + "logprob_reward": -0.11295662075281143, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295662075281143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13923279941082, + 0.07349938154220581, + 0.0892525240778923, + 0.045932602137327194, + 0.06875892728567123, + 0.10187959671020508, + 0.08740895241498947, + 0.23705245554447174, + 0.12657539546489716, + 0.030407369136810303 + ], + "credit_map": [ + 0.01971040479838848, + 0.00970130693167448, + 0.006726508028805256, + 0.010849729180335999, + 0.026109442114830017, + 0.017639417201280594, + 0.025571973994374275, + 0.022924011573195457, + 0.008777539245784283, + 0.00488860672339797, + 0.0033337227068841457, + 0.005283226724714041, + 0.010614650323987007, + 0.019600525498390198, + 0.011029030196368694, + 0.00997207686305046, + 0.013900860212743282, + 0.006043076049536467, + 0.004770135972648859, + 0.005893423687666655, + 0.009300279431045055, + 0.015718577429652214, + 0.012601478025317192, + 0.021024692803621292, + 0.005928738042712212, + 0.004111149813979864, + 0.003016287926584482, + 0.002926552202552557, + 0.0039362357929348946, + 0.007055284455418587, + 0.007024981547147036, + 0.011933373287320137, + 0.011554024182260036, + 0.00807128380984068, + 0.005949059501290321, + 0.006757405586540699, + 0.006854095030575991, + 0.010444706305861473, + 0.00957048125565052, + 0.009557871147990227, + 0.017533062025904655, + 0.00909501500427723, + 0.0070071774534881115, + 0.008292172104120255, + 0.011472085490822792, + 0.018514035269618034, + 0.014730970375239849, + 0.015235072001814842, + 0.00839022733271122, + 0.006719967816025019, + 0.0044157677330076694, + 0.009127920493483543, + 0.01734897308051586, + 0.016792235895991325, + 0.014612779021263123, + 0.010001084767282009, + 0.022464634850621223, + 0.013324047438800335, + 0.01149083860218525, + 0.013399261049926281, + 0.006387488916516304, + 0.07185564190149307, + 0.016390515491366386, + 0.08174002170562744, + 0.015065501444041729, + 0.01774463802576065, + 0.006572211626917124, + 0.008297295309603214, + 0.007690555416047573, + 0.03672601282596588, + 0.009959932416677475, + 0.024519238620996475, + 0.003533310955390334, + 0.002720598364248872, + 0.0019467484671622515, + 0.0024881421122699976, + 0.004627610091120005, + 0.00474328501150012, + 0.00544756418094039, + 0.004900109954178333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.11295662075281143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901675999164581 + ], + "opposite_logprob_tokens": [ + -2.57901668548584 + ], + "logprob_reward": -0.07901675999164581, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901675999164581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2721920907497406, + 0.13499590754508972, + 0.10710060596466064, + 0.13133865594863892, + 0.061962008476257324, + 0.029046474024653435, + 0.025578388944268227, + 0.041885800659656525, + 0.1046304702758789, + 0.0912695825099945 + ], + "credit_map": [ + 0.04114841669797897, + 0.024699075147509575, + 0.014669043011963367, + 0.01953682117164135, + 0.051845431327819824, + 0.039411935955286026, + 0.041504546999931335, + 0.039376817643642426, + 0.01905718632042408, + 0.011982198804616928, + 0.01095630880445242, + 0.016911158338189125, + 0.012004186399281025, + 0.02383401058614254, + 0.022648712620139122, + 0.01760215312242508, + 0.01820320636034012, + 0.013805256225168705, + 0.010061712004244328, + 0.009574256837368011, + 0.008746362291276455, + 0.01801409386098385, + 0.013806148432195187, + 0.014889566227793694, + 0.017085397616028786, + 0.014917793683707714, + 0.009851552546024323, + 0.01324909832328558, + 0.013963830657303333, + 0.02120806649327278, + 0.016881128773093224, + 0.02418178878724575, + 0.009088004007935524, + 0.008565962314605713, + 0.005857836455106735, + 0.006012920755892992, + 0.005530356429517269, + 0.008415929973125458, + 0.009428509511053562, + 0.009062491357326508, + 0.00489367451518774, + 0.0033989117946475744, + 0.0022426173090934753, + 0.002947355154901743, + 0.002419311087578535, + 0.005042759235948324, + 0.0034639870282262564, + 0.004637859761714935, + 0.003415903775021434, + 0.0020722339395433664, + 0.0011387491831555963, + 0.0017467985162511468, + 0.004106397740542889, + 0.005292197689414024, + 0.004047236870974302, + 0.003758871229365468, + 0.004487301222980022, + 0.0030125074554234743, + 0.0025167299900203943, + 0.0037994326557964087, + 0.003308504121378064, + 0.009174050763249397, + 0.006704720668494701, + 0.008882556110620499, + 0.024038761854171753, + 0.00890034157782793, + 0.00463097682222724, + 0.004032253287732601, + 0.004302775487303734, + 0.037239495664834976, + 0.00662616454064846, + 0.014859700575470924, + 0.01072816550731659, + 0.008790750056505203, + 0.006320164538919926, + 0.008491055108606815, + 0.013526508584618568, + 0.015850095078349113, + 0.012989417649805546, + 0.01457342691719532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07901675999164581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036202622577548027 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036202622577548027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036202622577548027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14164115488529205, + 0.06320255994796753, + 0.06529832631349564, + 0.1418580412864685, + 0.1032930463552475, + 0.12407085299491882, + 0.13279685378074646, + 0.0754161849617958, + 0.10640005767345428, + 0.04602298513054848 + ], + "credit_map": [ + 0.01732327789068222, + 0.01235514972358942, + 0.005853088106960058, + 0.008745158091187477, + 0.0328262560069561, + 0.023733064532279968, + 0.022015584632754326, + 0.01878957450389862, + 0.008755053393542767, + 0.004064368549734354, + 0.003800663398578763, + 0.00399013003334403, + 0.008182594552636147, + 0.017015479505062103, + 0.008029337041079998, + 0.00936493556946516, + 0.009569871239364147, + 0.005332491360604763, + 0.004338259808719158, + 0.005400523077696562, + 0.00792447105050087, + 0.011452157981693745, + 0.0110562052577734, + 0.01022434327751398, + 0.01778995990753174, + 0.011905037797987461, + 0.009263361804187298, + 0.014203101396560669, + 0.010300671681761742, + 0.021234294399619102, + 0.02314264513552189, + 0.03401895612478256, + 0.015538548119366169, + 0.010759166441857815, + 0.007262129336595535, + 0.009756798855960369, + 0.011423100717365742, + 0.01571730524301529, + 0.015025175176560879, + 0.017810821533203125, + 0.017036983743309975, + 0.007961597293615341, + 0.005346882622689009, + 0.007706653792411089, + 0.01873077265918255, + 0.0278910044580698, + 0.01645081304013729, + 0.02294614352285862, + 0.0144735686480999, + 0.009549948386847973, + 0.006037009414285421, + 0.01056874729692936, + 0.028293505311012268, + 0.02565346099436283, + 0.022574370726943016, + 0.01564624533057213, + 0.009084140881896019, + 0.006295644678175449, + 0.0036937729455530643, + 0.0033660028129816055, + 0.0047588408924639225, + 0.014975487254559994, + 0.00837442185729742, + 0.02486787736415863, + 0.022161632776260376, + 0.00820185150951147, + 0.00451102526858449, + 0.00576663576066494, + 0.004167781677097082, + 0.038718558847904205, + 0.006900177337229252, + 0.01597239077091217, + 0.00622826162725687, + 0.0034534246660768986, + 0.0035583116114139557, + 0.003992896061390638, + 0.0070581380277872086, + 0.008570310659706593, + 0.006310189608484507, + 0.006851450074464083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0036202622577548027 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot shot from a slightly overhead angle focuses the viewer's attention on a softly overhead-lit scene. The focal point of the image is a hand clad in a snug-fitting black lambskin glove, its surface gleaming with an understated yet elegant sheen, the leather's grain clearly visible in the light. The fingers of this hand are slightly bent, with only the index finger extended. With an extremely slow, appreciative rhythm, its fingertips carefully trace the surface of a brand-new tire. The tire, a purest matte black, is covered in an intricate and precise geometric tread pattern, each groove edge razor-sharp as if it were new. As the gloved fingertip glides across, the soft leather deforms minutely at the point of contact, creating a striking textural contrast with the tough rubber. The light precisely outlines the three-dimensional contours of the tire blocks and the subtle drainage grooves, casting soft shadows in the deep ravines, and blurring the background into an unrecognizable dark color, thus pushing this tactile and focused interaction between hand and tire into absolute visual focus, creating a tranquil atmosphere that is almost obsessive about the beauty of craftsmanship, performance and materials.", + "seed": 21815, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_11/index_8549_seed_21815_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": -0.07434283485636115, + "vlm_logprob": -0.07434283485636115, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.07434283485636115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044739481061697006 + ], + "opposite_logprob_tokens": [ + -3.4197394847869873 + ], + "logprob_reward": -0.044739481061697006, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.044739481061697006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15280921757221222, + 0.05211377516388893, + 0.09532347321510315, + 0.2413458526134491, + 0.1271316409111023, + 0.06048308312892914, + 0.06576845049858093, + 0.0598563589155674, + 0.034896284341812134, + 0.1102718710899353 + ], + "credit_map": [ + 0.03344224393367767, + 0.00968931708484888, + 0.00788915529847145, + 0.013889438472688198, + 0.014359204098582268, + 0.016420068219304085, + 0.026064163073897362, + 0.03105562925338745, + 0.007277332711964846, + 0.0035799359902739525, + 0.0036570229567587376, + 0.005838220939040184, + 0.005953535437583923, + 0.009205730631947517, + 0.010509027168154716, + 0.006092969793826342, + 0.012150895781815052, + 0.005325709003955126, + 0.005446910858154297, + 0.010186586529016495, + 0.006433392409235239, + 0.02064117044210434, + 0.021359246224164963, + 0.013779562897980213, + 0.038602426648139954, + 0.006735598668456078, + 0.005861667916178703, + 0.017862731590867043, + 0.010795284062623978, + 0.02123338170349598, + 0.09601209312677383, + 0.044242676347494125, + 0.01812848262488842, + 0.011829307302832603, + 0.012914873659610748, + 0.012083792127668858, + 0.01325288973748684, + 0.014738437719643116, + 0.02600272186100483, + 0.01818113774061203, + 0.011595268733799458, + 0.005756585393100977, + 0.004057705868035555, + 0.0062287598848342896, + 0.00436380784958601, + 0.0070085106417536736, + 0.01178137119859457, + 0.009691079147160053, + 0.0092874551191926, + 0.005994786508381367, + 0.004107627086341381, + 0.007525318767875433, + 0.007846294902265072, + 0.01227107085287571, + 0.00884231273084879, + 0.009893584996461868, + 0.006215259432792664, + 0.0033917841501533985, + 0.002279168926179409, + 0.005613123532384634, + 0.015873633325099945, + 0.007070670835673809, + 0.0052510881796479225, + 0.01416163519024849, + 0.00501440092921257, + 0.002599686151370406, + 0.002353379502892494, + 0.006458058021962643, + 0.0031166861299425364, + 0.0038063395768404007, + 0.004807079676538706, + 0.006740657612681389, + 0.01658807508647442, + 0.01153440959751606, + 0.00670420890673995, + 0.022345080971717834, + 0.009362485259771347, + 0.01120526809245348, + 0.013732047751545906, + 0.018800299614667892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.044739481061697006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06213880702853203 + ], + "opposite_logprob_tokens": [ + -2.812138795852661 + ], + "logprob_reward": -0.06213880702853203, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06213880702853203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37458205223083496, + 0.11138766258955002, + 0.056905101984739304, + 0.032012611627578735, + 0.1120375543832779, + 0.0836881473660469, + 0.05841252580285072, + 0.04165521264076233, + 0.05988924950361252, + 0.06942984461784363 + ], + "credit_map": [ + 0.07151367515325546, + 0.03199947625398636, + 0.027091294527053833, + 0.03442294895648956, + 0.036824584007263184, + 0.04401751607656479, + 0.059931740164756775, + 0.06878078728914261, + 0.01715237647294998, + 0.008516618981957436, + 0.00768641522154212, + 0.008920981548726559, + 0.012753591872751713, + 0.020787375047802925, + 0.02164774388074875, + 0.013922561891376972, + 0.010114353150129318, + 0.004247245844453573, + 0.0033784620463848114, + 0.006004301831126213, + 0.005181020125746727, + 0.011871147900819778, + 0.00960941519588232, + 0.0064991554245352745, + 0.007534878794103861, + 0.0020750320982187986, + 0.0017225902993232012, + 0.003991568461060524, + 0.0026709800586104393, + 0.0035755112767219543, + 0.0038187813479453325, + 0.006623271852731705, + 0.018985005095601082, + 0.011840277351439, + 0.012523218058049679, + 0.01413739938288927, + 0.0077059403993189335, + 0.01636919379234314, + 0.013662490993738174, + 0.016814028844237328, + 0.012605206109583378, + 0.0075347935780882835, + 0.007741484325379133, + 0.008827063255012035, + 0.005991935729980469, + 0.010456401854753494, + 0.01843741349875927, + 0.012093852274119854, + 0.0107793640345335, + 0.0056434813886880875, + 0.003949897829443216, + 0.0056512802839279175, + 0.006176939699798822, + 0.010121726430952549, + 0.008500528521835804, + 0.007589310873299837, + 0.0071531315334141254, + 0.0034046166110783815, + 0.0030344778206199408, + 0.005195518955588341, + 0.00549552496522665, + 0.005916227586567402, + 0.005871639586985111, + 0.0055840774439275265, + 0.008631509728729725, + 0.004536751192063093, + 0.003911146894097328, + 0.008677677251398563, + 0.005406566895544529, + 0.006256760098040104, + 0.009711911901831627, + 0.012756923213601112, + 0.010116406716406345, + 0.007441216614097357, + 0.005107842851430178, + 0.013839085586369038, + 0.006240911316126585, + 0.006446714513003826, + 0.008635671809315681, + 0.011601995676755905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.06213880702853203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003289744257927 + ], + "opposite_logprob_tokens": [ + -2.3503289222717285 + ], + "logprob_reward": -0.1003289744257927, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003289744257927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13257500529289246, + 0.06987784057855606, + 0.12120956182479858, + 0.1061343252658844, + 0.14312979578971863, + 0.05480588227510452, + 0.16506963968276978, + 0.09062894433736801, + 0.05252978578209877, + 0.06403926014900208 + ], + "credit_map": [ + 0.017826789990067482, + 0.015075511299073696, + 0.010576623491942883, + 0.015219729393720627, + 0.015629569068551064, + 0.020350929349660873, + 0.021026140078902245, + 0.016869693994522095, + 0.0097044100984931, + 0.005670960061252117, + 0.0046311537735164165, + 0.007775062695145607, + 0.00809864979237318, + 0.012176459655165672, + 0.010696902871131897, + 0.011124233715236187, + 0.022871389985084534, + 0.006751336622983217, + 0.005002994555979967, + 0.012248709797859192, + 0.006457487121224403, + 0.019150683656334877, + 0.02939995378255844, + 0.019326990470290184, + 0.015737207606434822, + 0.004394500516355038, + 0.003077107947319746, + 0.008172375150024891, + 0.004796146415174007, + 0.01007644273340702, + 0.03974628448486328, + 0.02013425901532173, + 0.021622460335493088, + 0.011193373240530491, + 0.011897302232682705, + 0.014874215237796307, + 0.013792107813060284, + 0.01731887273490429, + 0.030740173533558846, + 0.021691277623176575, + 0.01349355187267065, + 0.0039284201338887215, + 0.003485830035060644, + 0.004609543830156326, + 0.0035086411517113447, + 0.00640439335256815, + 0.009230549447238445, + 0.010144946165382862, + 0.028982214629650116, + 0.018124673515558243, + 0.010679385624825954, + 0.016482267528772354, + 0.018151871860027313, + 0.02369668520987034, + 0.024990497156977654, + 0.023962020874023438, + 0.013916603289544582, + 0.005731088574975729, + 0.003926461562514305, + 0.011862700805068016, + 0.021916238591074944, + 0.014004719443619251, + 0.008012122474610806, + 0.01125900074839592, + 0.00784376822412014, + 0.0036971126683056355, + 0.0030242111533880234, + 0.007509132847189903, + 0.004764004610478878, + 0.005334567744284868, + 0.008431206457316875, + 0.011925775557756424, + 0.011290512047708035, + 0.005213739350438118, + 0.006438697688281536, + 0.008400571532547474, + 0.007666323333978653, + 0.005516474600881338, + 0.006981252692639828, + 0.012531682848930359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.1003289744257927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604061871767044 + ], + "opposite_logprob_tokens": [ + -1.910406231880188 + ], + "logprob_reward": -0.1604061871767044, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604061871767044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10994050651788712, + 0.10228406637907028, + 0.06283144652843475, + 0.150366872549057, + 0.3150162696838379, + 0.04695642739534378, + 0.05357064679265022, + 0.04316624626517296, + 0.02562507428228855, + 0.09024246037006378 + ], + "credit_map": [ + 0.01754545420408249, + 0.007400395814329386, + 0.00853852927684784, + 0.012981613166630268, + 0.009377749636769295, + 0.014897288754582405, + 0.014980236068367958, + 0.024219226092100143, + 0.01901998370885849, + 0.006837841589003801, + 0.006424452643841505, + 0.007272583898156881, + 0.012283327989280224, + 0.015497125685214996, + 0.021572891622781754, + 0.013375842943787575, + 0.008800198324024677, + 0.00413289200514555, + 0.003236598102375865, + 0.005599462892860174, + 0.0045601241290569305, + 0.012399972416460514, + 0.011563773266971111, + 0.012538420967757702, + 0.02410576120018959, + 0.0047785453498363495, + 0.004565748386085033, + 0.011391867883503437, + 0.007238736841827631, + 0.010380584746599197, + 0.059376925230026245, + 0.028528695926070213, + 0.04366300627589226, + 0.032925911247730255, + 0.036252461373806, + 0.039918847382068634, + 0.024730602279305458, + 0.03463265299797058, + 0.05402711033821106, + 0.0488656647503376, + 0.010536259040236473, + 0.0033527067862451077, + 0.003479287726804614, + 0.004388406407088041, + 0.0025059001054614782, + 0.004830564372241497, + 0.009476428851485252, + 0.008386867120862007, + 0.008938773535192013, + 0.0063085309229791164, + 0.003450799500569701, + 0.006207129918038845, + 0.006702568847686052, + 0.006967531051486731, + 0.007261048071086407, + 0.007734254933893681, + 0.005892810877412558, + 0.002583710942417383, + 0.001861531171016395, + 0.004646332003176212, + 0.009803949855268002, + 0.004652121104300022, + 0.003580260556191206, + 0.010145524516701698, + 0.004570977762341499, + 0.0017726441146805882, + 0.00232231174595654, + 0.0036865798756480217, + 0.002134832553565502, + 0.0023267276119440794, + 0.00294442824088037, + 0.005866568069905043, + 0.01536789070814848, + 0.008865270763635635, + 0.009225890040397644, + 0.012584016658365726, + 0.00914326123893261, + 0.008287178352475166, + 0.010855547152459621, + 0.015913400799036026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.1604061871767044 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 with a static camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand wearing a snug-fitting black lambskin glove with visible leather grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the fingers slightly bent with only the index finger extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the object a brand-new matte black tire with a precise geometric tread pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the fingertip trace the tire surface while the background is blurred into an unrecognizable dark color?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041007245890796185 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.0041007245890796185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041007245890796185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17589904367923737, + 0.10871250182390213, + 0.12190870940685272, + 0.08016679435968399, + 0.10465647280216217, + 0.10452377051115036, + 0.06777176260948181, + 0.12665055692195892, + 0.038916852325201035, + 0.0707935243844986 + ], + "credit_map": [ + 0.027278104797005653, + 0.009895755909383297, + 0.016560550779104233, + 0.016457397490739822, + 0.017467480152845383, + 0.02275642193853855, + 0.026915526017546654, + 0.038567814975976944, + 0.019048916175961494, + 0.008955716155469418, + 0.008228005841374397, + 0.009360485710203648, + 0.01270601712167263, + 0.015039944089949131, + 0.022785328328609467, + 0.01258808746933937, + 0.02042205259203911, + 0.009839487262070179, + 0.007942848838865757, + 0.013040953315794468, + 0.009215639904141426, + 0.015173852443695068, + 0.02450493723154068, + 0.021768935024738312, + 0.013667459599673748, + 0.006907661445438862, + 0.006345713045448065, + 0.012593804858624935, + 0.0068199485540390015, + 0.008565029129385948, + 0.011076876893639565, + 0.014190295711159706, + 0.013350696302950382, + 0.010634667240083218, + 0.009639900177717209, + 0.010679280385375023, + 0.010330942459404469, + 0.015249587595462799, + 0.018788298591971397, + 0.01598309725522995, + 0.017244936898350716, + 0.01076019648462534, + 0.007298223674297333, + 0.011449524201452732, + 0.0074662393890321255, + 0.01300876960158348, + 0.022011008113622665, + 0.015284867957234383, + 0.010502539575099945, + 0.0056694988161325455, + 0.004816439934074879, + 0.007135866675525904, + 0.006424737628549337, + 0.014057492837309837, + 0.009458750486373901, + 0.009706433862447739, + 0.018775606527924538, + 0.007716482970863581, + 0.007805361412465572, + 0.01861543580889702, + 0.00722062261775136, + 0.01991276815533638, + 0.014606044627726078, + 0.03199823573231697, + 0.007235633209347725, + 0.00271895551122725, + 0.002581983804702759, + 0.004475898575037718, + 0.0037076633889228106, + 0.00398830184713006, + 0.0033520457800477743, + 0.01085637416690588, + 0.011264112778007984, + 0.011215964332222939, + 0.005132347345352173, + 0.009756727144122124, + 0.005995023995637894, + 0.005818684585392475, + 0.0075285779312253, + 0.014082091860473156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0041007245890796185 + } + ], + "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/0000110/group_12/index_tensor([6728, 6728])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_12/index_tensor([6728, 6728])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11a99da31c9e6b4ef5ab690811541ad80a2e3d66 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_12/index_tensor([6728, 6728])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot captures a male athlete's climactic moment, his powerful physique clad in a sleeveless royal blue leotard as he executes a graceful sprint from the gleaming, polished steel parallel bars. With short, black hair damp from exertion and hands coated in white chalk, he thrusts forward decisively, his body spinning in a blur of controlled motion through the air. He lands with pinpoint precision on a thick, navy blue mat, his feet striking the surface simultaneously with a solid, satisfying thud. His knees bend deeply to absorb the force, his arms briefly extended for balance, before he straightens into a resolute, triumphant pose, his expression filled with an intense, focused sense of accomplishment. As he holds his pose, the camera begins a steady, continuous crane shot, tilting upward and pulling back to reveal a vast, well-equipped, modern gymnasium. Bright, diffuse overhead lighting illuminates the space, and the other state-of-the-art equipment gleaming in the background—the pommel horse, rings, and a sprawling floor exercise area—establishes an atmosphere of elite athleticism and disciplinary perfection within the vast, professional training hall.", + "seed": 3946, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_12/index_6728_seed_3946_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": -0.0588323812931776, + "vlm_logprob": -0.0588323812931776, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.0588323812931776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08473819494247437 + ], + "opposite_logprob_tokens": [ + -2.584738254547119 + ], + "logprob_reward": -0.08473819494247437, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08473819494247437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0886685773730278, + 0.05398460477590561, + 0.06345214694738388, + 0.21141959726810455, + 0.07303912192583084, + 0.04733700305223465, + 0.12292101234197617, + 0.0959654375910759, + 0.22507470846176147, + 0.018137764185667038 + ], + "credit_map": [ + 0.007551205810159445, + 0.010320362634956837, + 0.013205960392951965, + 0.011849119327962399, + 0.010363717563450336, + 0.01084858737885952, + 0.01521309744566679, + 0.009316535666584969, + 0.005908419378101826, + 0.007194153964519501, + 0.009542012587189674, + 0.007435585372149944, + 0.005869470536708832, + 0.005982291419059038, + 0.0069671375676989555, + 0.005085536744445562, + 0.004403700586408377, + 0.007156155537813902, + 0.011956377886235714, + 0.013085330836474895, + 0.0034165075048804283, + 0.007395741064101458, + 0.007395111955702305, + 0.008643229492008686, + 0.05040861666202545, + 0.0233469195663929, + 0.03679199516773224, + 0.012755884788930416, + 0.023139391094446182, + 0.017989329993724823, + 0.016584651544690132, + 0.03040282242000103, + 0.01209208369255066, + 0.010010198689997196, + 0.0069283246994018555, + 0.007608313113451004, + 0.006425155326724052, + 0.01394052803516388, + 0.006259204354137182, + 0.00977531261742115, + 0.005824149586260319, + 0.008423087187111378, + 0.006352574098855257, + 0.005552881397306919, + 0.004636803641915321, + 0.0046801939606666565, + 0.005688526201993227, + 0.006178791634738445, + 0.01833510398864746, + 0.014902183786034584, + 0.02221417799592018, + 0.009513521566987038, + 0.012171308510005474, + 0.0226316899061203, + 0.008368836715817451, + 0.01478420477360487, + 0.009958862327039242, + 0.011348831467330456, + 0.011014288291335106, + 0.013543549925088882, + 0.00957898423075676, + 0.020201098173856735, + 0.00945674441754818, + 0.010863091796636581, + 0.013011795468628407, + 0.026595793664455414, + 0.03412958234548569, + 0.033528558909893036, + 0.02261543832719326, + 0.024090593680739403, + 0.022790830582380295, + 0.048312149941921234, + 0.002320462604984641, + 0.0023575355298817158, + 0.0020970096811652184, + 0.0014247677754610777, + 0.004203819669783115, + 0.0021739418152719736, + 0.0016763554885983467, + 0.001883873832412064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.08473819494247437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004771381616592 + ], + "opposite_logprob_tokens": [ + -2.695047616958618 + ], + "logprob_reward": -0.07004771381616592, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004771381616592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09234227985143661, + 0.09416275471448898, + 0.12286176532506943, + 0.1745913028717041, + 0.09772909432649612, + 0.1727515608072281, + 0.09640670567750931, + 0.05610153079032898, + 0.04528392478823662, + 0.047769103199243546 + ], + "credit_map": [ + 0.013463041745126247, + 0.014005390927195549, + 0.014379664324223995, + 0.014369831420481205, + 0.008067644201219082, + 0.010510957799851894, + 0.009541137143969536, + 0.008004616014659405, + 0.007676930632442236, + 0.006440614350140095, + 0.013962890952825546, + 0.014640701934695244, + 0.012358385138213634, + 0.01175144873559475, + 0.015307565219700336, + 0.01202421449124813, + 0.013525264337658882, + 0.010475785471498966, + 0.03213009238243103, + 0.014782522805035114, + 0.009834608994424343, + 0.009562884457409382, + 0.017335588112473488, + 0.015215015038847923, + 0.050823237746953964, + 0.022463800385594368, + 0.023731982335448265, + 0.00965531263500452, + 0.013878938741981983, + 0.020028704777359962, + 0.0108214495703578, + 0.02318788692355156, + 0.010773403570055962, + 0.009659108705818653, + 0.02257378026843071, + 0.009119667112827301, + 0.006115520838648081, + 0.009157448075711727, + 0.00563597260043025, + 0.02469419129192829, + 0.022132383659482002, + 0.011932704597711563, + 0.021947667002677917, + 0.007831699214875698, + 0.03607785701751709, + 0.039411790668964386, + 0.010392929427325726, + 0.023024532943964005, + 0.010549512691795826, + 0.005541108548641205, + 0.025708766654133797, + 0.003948915284126997, + 0.007226413115859032, + 0.010172692127525806, + 0.00423944927752018, + 0.029019853100180626, + 0.00938070472329855, + 0.00610360037535429, + 0.006476286798715591, + 0.005231130868196487, + 0.004674479831010103, + 0.013134729117155075, + 0.004210528451949358, + 0.006890072021633387, + 0.005766578018665314, + 0.00686822971329093, + 0.006116651929914951, + 0.004262919537723064, + 0.0031568617559969425, + 0.005981354042887688, + 0.002995810005813837, + 0.010135523043572903, + 0.006907588802278042, + 0.005410507787019014, + 0.005640395916998386, + 0.0041311755776405334, + 0.0049225869588553905, + 0.009576422162353992, + 0.002854840597137809, + 0.008325586095452309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.07004771381616592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043056778609752655 + ], + "opposite_logprob_tokens": [ + -3.1680567264556885 + ], + "logprob_reward": -0.043056778609752655, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043056778609752655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07509949058294296, + 0.053464990109205246, + 0.15093959867954254, + 0.33579131960868835, + 0.12238273024559021, + 0.1348467916250229, + 0.039395421743392944, + 0.022137532010674477, + 0.04095190763473511, + 0.024990225210785866 + ], + "credit_map": [ + 0.011536956764757633, + 0.017688635736703873, + 0.008081686682999134, + 0.011343291029334068, + 0.007040564436465502, + 0.007211492862552404, + 0.006752028129994869, + 0.005444823764264584, + 0.0057403999380767345, + 0.005107593722641468, + 0.007401054259389639, + 0.008255324326455593, + 0.006088497117161751, + 0.006464064121246338, + 0.007740113418549299, + 0.0066679371520876884, + 0.011298291385173798, + 0.016165487468242645, + 0.031970251351594925, + 0.025607870891690254, + 0.008883312344551086, + 0.014516511000692844, + 0.01928158849477768, + 0.023216284811496735, + 0.10075797140598297, + 0.04510060325264931, + 0.04888734593987465, + 0.018206892535090446, + 0.021313898265361786, + 0.03503793850541115, + 0.020663434639573097, + 0.04582320153713226, + 0.015433333814144135, + 0.020744293928146362, + 0.012412783689796925, + 0.014654725790023804, + 0.011899477802217007, + 0.0202623438090086, + 0.010246405377984047, + 0.016729356721043587, + 0.016413908451795578, + 0.013822942040860653, + 0.02246215008199215, + 0.012086649425327778, + 0.014955345541238785, + 0.021975111216306686, + 0.011303616687655449, + 0.021827049553394318, + 0.004698572214692831, + 0.004708573222160339, + 0.005219681654125452, + 0.005091572646051645, + 0.0055894735269248486, + 0.004854191094636917, + 0.0036102801095694304, + 0.005623071454465389, + 0.0033144797198474407, + 0.0026279648300260305, + 0.002235384425148368, + 0.0027257392648607492, + 0.0022446028888225555, + 0.003876456990838051, + 0.0021654816810041666, + 0.0029474191833287477, + 0.004367605317384005, + 0.00418365141376853, + 0.005706388968974352, + 0.004986260086297989, + 0.0050981249660253525, + 0.005301326978951693, + 0.00537773547694087, + 0.005930809769779444, + 0.0035252352245151997, + 0.002681967569515109, + 0.0030547974165529013, + 0.0028833006508648396, + 0.002555929124355316, + 0.0041835554875433445, + 0.0027696865145117044, + 0.003335752757266164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.043056778609752655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999125331640244 + ], + "opposite_logprob_tokens": [ + -2.69499135017395 + ], + "logprob_reward": -0.06999125331640244, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999125331640244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06826990842819214, + 0.06140416860580444, + 0.04960060492157936, + 0.16654445230960846, + 0.12297946214675903, + 0.20121926069259644, + 0.0253708828240633, + 0.05294322222471237, + 0.21333132684230804, + 0.03833670914173126 + ], + "credit_map": [ + 0.008758570067584515, + 0.011961573734879494, + 0.008401130326092243, + 0.008992013521492481, + 0.009579936042428017, + 0.008103185333311558, + 0.006362146232277155, + 0.006111351307481527, + 0.006774664856493473, + 0.007576158735901117, + 0.009496685117483139, + 0.009396800771355629, + 0.005869836080819368, + 0.0058230687864124775, + 0.008433292619884014, + 0.008033661171793938, + 0.005161042790859938, + 0.004127048887312412, + 0.007811123970896006, + 0.010377789847552776, + 0.004463387653231621, + 0.00609159329906106, + 0.006355388090014458, + 0.005213229916989803, + 0.0486002154648304, + 0.022625429555773735, + 0.025248918682336807, + 0.009604016318917274, + 0.015056032687425613, + 0.015300252474844456, + 0.010751648806035519, + 0.019357947632670403, + 0.019736602902412415, + 0.02057170309126377, + 0.017248686403036118, + 0.012521148659288883, + 0.010172990150749683, + 0.014584059827029705, + 0.008343067951500416, + 0.019801201298832893, + 0.025951722636818886, + 0.01615181751549244, + 0.021022414788603783, + 0.010888359509408474, + 0.047064658254384995, + 0.043558213859796524, + 0.014721279963850975, + 0.021860798820853233, + 0.004094325937330723, + 0.0024043044541031122, + 0.0034838595893234015, + 0.001915664877742529, + 0.003043833654373884, + 0.004743737634271383, + 0.0024401189293712378, + 0.003245037514716387, + 0.007186106871813536, + 0.006449172738939524, + 0.006528975907713175, + 0.004186590667814016, + 0.0052568246610462666, + 0.011727566830813885, + 0.0036450079642236233, + 0.007962977513670921, + 0.019219055771827698, + 0.024531139060854912, + 0.03195228800177574, + 0.028555266559123993, + 0.012106949463486671, + 0.023965056985616684, + 0.018085559830069542, + 0.05491601303219795, + 0.006231484469026327, + 0.0046223909594118595, + 0.004769263789057732, + 0.003119375789538026, + 0.00581815792247653, + 0.005824225023388863, + 0.0030740932561457157, + 0.004877718165516853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.06999125331640244 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026327965781092644 + ], + "opposite_logprob_tokens": [ + -3.6513278484344482 + ], + "logprob_reward": -0.026327965781092644, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026327965781092644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16910836100578308, + 0.09511353075504303, + 0.07834004610776901, + 0.17066940665245056, + 0.09973668307065964, + 0.2112458050251007, + 0.05674624443054199, + 0.026809344068169594, + 0.06658537685871124, + 0.025645168498158455 + ], + "credit_map": [ + 0.016896763816475868, + 0.03109358623623848, + 0.02222287654876709, + 0.033553604036569595, + 0.016469473019242287, + 0.012615850195288658, + 0.021706001833081245, + 0.014550223015248775, + 0.00993443001061678, + 0.011533671990036964, + 0.013857444748282433, + 0.01436098013073206, + 0.014464432373642921, + 0.0104772774502635, + 0.012522884644567966, + 0.007962421514093876, + 0.009171145036816597, + 0.006594114471226931, + 0.01591174118220806, + 0.01081862673163414, + 0.005471453070640564, + 0.006090183276683092, + 0.0100681371986866, + 0.014214659109711647, + 0.03875342011451721, + 0.01717647723853588, + 0.0298781618475914, + 0.01385191548615694, + 0.016569599509239197, + 0.014702354557812214, + 0.013061023317277431, + 0.026676472276449203, + 0.019202016294002533, + 0.013390684500336647, + 0.011531961150467396, + 0.010615955106914043, + 0.010324707254767418, + 0.014538859948515892, + 0.007518783211708069, + 0.012613723054528236, + 0.029752830043435097, + 0.012770580127835274, + 0.02888748236000538, + 0.010505734011530876, + 0.03925632685422897, + 0.04675331711769104, + 0.011194141581654549, + 0.032125409692525864, + 0.007575389929115772, + 0.005102707538753748, + 0.014971933327615261, + 0.0025935538578778505, + 0.004502051509916782, + 0.007811852730810642, + 0.002641861792653799, + 0.01154689583927393, + 0.004646537825465202, + 0.0031962504144757986, + 0.0027165606152266264, + 0.0037281052209436893, + 0.002465021563693881, + 0.004798554815351963, + 0.0021741786040365696, + 0.0030841371044516563, + 0.007601154502481222, + 0.007122257724404335, + 0.009723754599690437, + 0.005165880545973778, + 0.009168678894639015, + 0.011044347658753395, + 0.007862848229706287, + 0.00889645330607891, + 0.003482133848592639, + 0.002720224903896451, + 0.003094673855230212, + 0.001938051893375814, + 0.0031984541565179825, + 0.0033823021221905947, + 0.002321420470252633, + 0.005507909692823887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.026327965781092644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures a male athlete's climactic moment, his powerful physique clad in a sleeveless royal blue leotard as he executes a graceful sprint from the gleaming, polished steel parallel bars. With short, black hair damp from exertion and hands coated in white chalk, he thrusts forward decisively, his body spinning in a blur of controlled motion through the air. He lands with pinpoint precision on a thick, navy blue mat, his feet striking the surface simultaneously with a solid, satisfying thud. His knees bend deeply to absorb the force, his arms briefly extended for balance, before he straightens into a resolute, triumphant pose, his expression filled with an intense, focused sense of accomplishment. As he holds his pose, the camera begins a steady, continuous crane shot, tilting upward and pulling back to reveal a vast, well-equipped, modern gymnasium. Bright, diffuse overhead lighting illuminates the space, and the other state-of-the-art equipment gleaming in the background—the pommel horse, rings, and a sprawling floor exercise area—establishes an atmosphere of elite athleticism and disciplinary perfection within the vast, professional training hall.", + "seed": 3947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_12/index_6728_seed_3947_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": -0.10504882633686066, + "vlm_logprob": -0.10504882633686066, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.10504882633686066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20929855108261108 + ], + "opposite_logprob_tokens": [ + -1.7092986106872559 + ], + "logprob_reward": -0.20929855108261108, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20929855108261108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06546447426080704, + 0.06222071871161461, + 0.013343129307031631, + 0.01882389560341835, + 0.03141043707728386, + 0.031452544033527374, + 0.0305950790643692, + 0.09738176316022873, + 0.5706517696380615, + 0.07865617424249649 + ], + "credit_map": [ + 0.007740583270788193, + 0.0074698529206216335, + 0.005124329123646021, + 0.007993473671376705, + 0.010283833369612694, + 0.007239402271807194, + 0.01305657159537077, + 0.006556432694196701, + 0.007918475195765495, + 0.00978208426386118, + 0.004992039408534765, + 0.009258299134671688, + 0.009451033547520638, + 0.007328191306442022, + 0.006782165728509426, + 0.006708435248583555, + 0.0017773531144484878, + 0.0019148706924170256, + 0.001379080000333488, + 0.001572329318150878, + 0.002286792267113924, + 0.0017784537049010396, + 0.001031417865306139, + 0.0016028337413445115, + 0.001771572045981884, + 0.002780101727694273, + 0.002045049797743559, + 0.0018159262835979462, + 0.0038789897225797176, + 0.002274517435580492, + 0.0023835364263504744, + 0.0018742027459666133, + 0.0036617331206798553, + 0.004792220424860716, + 0.004968936089426279, + 0.0030476662795990705, + 0.0040916455909609795, + 0.0037163812667131424, + 0.0037063315976411104, + 0.0034255264326930046, + 0.004986837040632963, + 0.006049526855349541, + 0.0026878253556787968, + 0.003043856704607606, + 0.003207450732588768, + 0.004098156001418829, + 0.004226623103022575, + 0.0031522728968411684, + 0.003705480834469199, + 0.004566890187561512, + 0.003003182588145137, + 0.0026597667019814253, + 0.004520040471106768, + 0.00396370654925704, + 0.004046844784170389, + 0.004129169508814812, + 0.027297906577587128, + 0.006701166275888681, + 0.008199443109333515, + 0.016868973150849342, + 0.006828004494309425, + 0.013825401663780212, + 0.008359033614397049, + 0.009301832877099514, + 0.14813509583473206, + 0.03977033868432045, + 0.03215952217578888, + 0.07603949308395386, + 0.030703557655215263, + 0.15656760334968567, + 0.04749515652656555, + 0.039781052619218826, + 0.008451241068542004, + 0.012499360367655754, + 0.003600290510803461, + 0.007106449455022812, + 0.011210833676159382, + 0.018603378906846046, + 0.011807598173618317, + 0.005377026740461588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.20929855108261108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034245252609253 + ], + "opposite_logprob_tokens": [ + -2.3503425121307373 + ], + "logprob_reward": -0.10034245252609253, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034245252609253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10525967925786972, + 0.05280325934290886, + 0.06759557127952576, + 0.0261413361877203, + 0.0363132618367672, + 0.041120901703834534, + 0.035251516848802567, + 0.18584035336971283, + 0.43263331055641174, + 0.017040826380252838 + ], + "credit_map": [ + 0.01423230953514576, + 0.009926657192409039, + 0.011357633396983147, + 0.007484120782464743, + 0.014568768441677094, + 0.017635254189372063, + 0.01639101281762123, + 0.013663919642567635, + 0.006375575438141823, + 0.004886125214397907, + 0.007622601464390755, + 0.004627884831279516, + 0.008344040252268314, + 0.009189056232571602, + 0.00570798572152853, + 0.006049989722669125, + 0.007066283840686083, + 0.00867433100938797, + 0.006500673945993185, + 0.004913557320833206, + 0.0054618725553154945, + 0.010213506408035755, + 0.007399243302643299, + 0.01736610196530819, + 0.0044824955984950066, + 0.005105401389300823, + 0.0036547826603055, + 0.0015348708257079124, + 0.0031615227926522493, + 0.00259001855738461, + 0.0018822423880919814, + 0.003730000462383032, + 0.00566343916580081, + 0.0071535659953951836, + 0.0066511426120996475, + 0.0023662440944463015, + 0.0044075725600123405, + 0.004779567942023277, + 0.0025857838336378336, + 0.002705946331843734, + 0.005570668261498213, + 0.007208752445876598, + 0.0044959476217627525, + 0.0037624535616487265, + 0.004611251875758171, + 0.005213786382228136, + 0.004426952451467514, + 0.005831088870763779, + 0.0040156906470656395, + 0.005643852055072784, + 0.0027293232269585133, + 0.004734260495752096, + 0.004430343396961689, + 0.004692608490586281, + 0.00518405344337225, + 0.0038213832303881645, + 0.04462708532810211, + 0.016747621819376945, + 0.02238212525844574, + 0.046416912227869034, + 0.010952973738312721, + 0.024810710921883583, + 0.011909101158380508, + 0.007993821054697037, + 0.11360780149698257, + 0.03004307858645916, + 0.026333708316087723, + 0.05682947859168053, + 0.02222723700106144, + 0.11800996214151382, + 0.035316433757543564, + 0.030265644192695618, + 0.0027200730983167887, + 0.002637255471199751, + 0.0017813737504184246, + 0.0014558536931872368, + 0.0024360434617847204, + 0.002460598247125745, + 0.001974300714209676, + 0.001575328060425818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.10034245252609253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305928945541382 + ], + "opposite_logprob_tokens": [ + -3.1680593490600586 + ], + "logprob_reward": -0.04305928945541382, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305928945541382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053545258939266205, + 0.08211372792720795, + 0.04218403995037079, + 0.06252089142799377, + 0.0899401307106018, + 0.17932960391044617, + 0.06392098218202591, + 0.15877333283424377, + 0.21992748975753784, + 0.04774460569024086 + ], + "credit_map": [ + 0.007024340331554413, + 0.008957496844232082, + 0.005814605858176947, + 0.006581650581210852, + 0.007218418177217245, + 0.006775810848921537, + 0.006422741804271936, + 0.004750192165374756, + 0.01163895707577467, + 0.009857668541371822, + 0.008862324059009552, + 0.008470535278320312, + 0.012531526386737823, + 0.01170459482818842, + 0.008877713233232498, + 0.01017040852457285, + 0.007444049697369337, + 0.005459855310618877, + 0.0055339401587843895, + 0.003582230070605874, + 0.007501148618757725, + 0.005570492707192898, + 0.0029083576519042253, + 0.004183965269476175, + 0.007923057302832603, + 0.010438804514706135, + 0.01003201398998499, + 0.004323465749621391, + 0.007079573813825846, + 0.006864884868264198, + 0.005249768495559692, + 0.010609324090182781, + 0.013174857944250107, + 0.01139914058148861, + 0.01796197146177292, + 0.009678300470113754, + 0.00996250007301569, + 0.00980174820870161, + 0.009750845842063427, + 0.008210765197873116, + 0.02035239152610302, + 0.04029928520321846, + 0.02451828308403492, + 0.01501700934022665, + 0.017260810360312462, + 0.025757042691111565, + 0.013514009304344654, + 0.02261076122522354, + 0.009877491742372513, + 0.006644216366112232, + 0.002622934291139245, + 0.004593366757035255, + 0.009206842631101608, + 0.015303212217986584, + 0.00908992625772953, + 0.006582989357411861, + 0.026038890704512596, + 0.012677891179919243, + 0.014004409313201904, + 0.050572168081998825, + 0.007231602445244789, + 0.018041277304291725, + 0.008988884277641773, + 0.02121821418404579, + 0.057380445301532745, + 0.015221371315419674, + 0.01228635199368, + 0.02768503874540329, + 0.011853507719933987, + 0.062438540160655975, + 0.017240500077605247, + 0.015821725130081177, + 0.008219446055591106, + 0.008399770595133305, + 0.0029016039334237576, + 0.002136766677722335, + 0.004051743075251579, + 0.01640286110341549, + 0.002882544184103608, + 0.002749871462583542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.04305928945541382 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14276055991649628 + ], + "opposite_logprob_tokens": [ + -2.0177605152130127 + ], + "logprob_reward": -0.14276055991649628, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14276055991649628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05391227453947067, + 0.060418542474508286, + 0.09881408512592316, + 0.08727218210697174, + 0.056065481156110764, + 0.047178272157907486, + 0.028572697192430496, + 0.32922133803367615, + 0.1986135095357895, + 0.03993157297372818 + ], + "credit_map": [ + 0.0076204342767596245, + 0.006560767535120249, + 0.007031692191958427, + 0.0039026234298944473, + 0.008722852915525436, + 0.007118544541299343, + 0.007415437139570713, + 0.005539925768971443, + 0.007932931184768677, + 0.00532235112041235, + 0.007615582551807165, + 0.010077361948788166, + 0.00791920442134142, + 0.009581610560417175, + 0.0046675181947648525, + 0.007301985751837492, + 0.015501860529184341, + 0.008822319097816944, + 0.010159787721931934, + 0.005731542594730854, + 0.011912945657968521, + 0.01400426309555769, + 0.012355992570519447, + 0.020325375720858574, + 0.015621473081409931, + 0.01631264202296734, + 0.011460970155894756, + 0.0061733415350317955, + 0.008960165083408356, + 0.009259890764951706, + 0.006066114641726017, + 0.013417582027614117, + 0.009202633053064346, + 0.006725893821567297, + 0.009014814160764217, + 0.008162381127476692, + 0.005258666817098856, + 0.006494821514934301, + 0.005326270125806332, + 0.005880000069737434, + 0.006482560187578201, + 0.01196636725217104, + 0.0033425292931497097, + 0.004654790740460157, + 0.005205495283007622, + 0.0065755825489759445, + 0.00502661382779479, + 0.0039243316277861595, + 0.0048407637514173985, + 0.004767929669469595, + 0.0022893687710165977, + 0.0023044084664434195, + 0.0038517361972481012, + 0.004305029287934303, + 0.002763445256277919, + 0.003450015326961875, + 0.055733516812324524, + 0.017368575558066368, + 0.023979438468813896, + 0.11519206315279007, + 0.00610852288082242, + 0.03763741999864578, + 0.01727888733148575, + 0.055922918021678925, + 0.06205081567168236, + 0.010830123908817768, + 0.00954510923475027, + 0.0172293558716774, + 0.008325653150677681, + 0.06508556753396988, + 0.011899750679731369, + 0.013647140003740788, + 0.004519665148109198, + 0.009177003987133503, + 0.0023928857408463955, + 0.0022954766172915697, + 0.004596487153321505, + 0.011087935417890549, + 0.0030913567170500755, + 0.0027707580011337996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.14276055991649628 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783278703689575 + ], + "opposite_logprob_tokens": [ + -3.529783248901367 + ], + "logprob_reward": -0.029783278703689575, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783278703689575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0812298133969307, + 0.04880622774362564, + 0.027055352926254272, + 0.03467358276247978, + 0.10930881649255753, + 0.039631281048059464, + 0.06160453334450722, + 0.192967027425766, + 0.2384917289018631, + 0.1662316769361496 + ], + "credit_map": [ + 0.010573035106062889, + 0.01121552661061287, + 0.006921112071722746, + 0.007623760960996151, + 0.012691662646830082, + 0.00845690444111824, + 0.015703020617365837, + 0.008044782094657421, + 0.0056205508299171925, + 0.004968773573637009, + 0.0060458071529865265, + 0.0061513218097388744, + 0.006322886329144239, + 0.006509412080049515, + 0.005832029972225428, + 0.007355441804975271, + 0.0035831534769386053, + 0.0030751838348805904, + 0.003335028188303113, + 0.0030821904074400663, + 0.0035678839776664972, + 0.004067460540682077, + 0.002263398375362158, + 0.0040810503996908665, + 0.004208122845739126, + 0.0052025699988007545, + 0.005356644280254841, + 0.002574724378064275, + 0.0038710811641067266, + 0.004010880831629038, + 0.003548672189936042, + 0.005900886841118336, + 0.013572835363447666, + 0.01731165684759617, + 0.01682772859930992, + 0.01172614935785532, + 0.014180234633386135, + 0.013449820689857006, + 0.011853811331093311, + 0.010386570356786251, + 0.0059870462864637375, + 0.011365666054189205, + 0.002686886815354228, + 0.002392880618572235, + 0.004245825111865997, + 0.006532858591526747, + 0.0027436872478574514, + 0.003676428459584713, + 0.011317552998661995, + 0.009735167026519775, + 0.004595027770847082, + 0.0054748826660215855, + 0.006792398635298014, + 0.009284933097660542, + 0.0066373469308018684, + 0.007767216768115759, + 0.04709929600358009, + 0.00976739265024662, + 0.016237176954746246, + 0.027430705726146698, + 0.00676187127828598, + 0.02047186717391014, + 0.007291280198842287, + 0.05790741741657257, + 0.05046238377690315, + 0.021082229912281036, + 0.015700573101639748, + 0.039268914610147476, + 0.017329758033156395, + 0.051185060292482376, + 0.02557327039539814, + 0.01788952201604843, + 0.03302386775612831, + 0.046918995678424835, + 0.01124302763491869, + 0.0064333351328969, + 0.01794837974011898, + 0.026268314570188522, + 0.010989796370267868, + 0.013405939564108849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.029783278703689575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures a male athlete's climactic moment, his powerful physique clad in a sleeveless royal blue leotard as he executes a graceful sprint from the gleaming, polished steel parallel bars. With short, black hair damp from exertion and hands coated in white chalk, he thrusts forward decisively, his body spinning in a blur of controlled motion through the air. He lands with pinpoint precision on a thick, navy blue mat, his feet striking the surface simultaneously with a solid, satisfying thud. His knees bend deeply to absorb the force, his arms briefly extended for balance, before he straightens into a resolute, triumphant pose, his expression filled with an intense, focused sense of accomplishment. As he holds his pose, the camera begins a steady, continuous crane shot, tilting upward and pulling back to reveal a vast, well-equipped, modern gymnasium. Bright, diffuse overhead lighting illuminates the space, and the other state-of-the-art equipment gleaming in the background—the pommel horse, rings, and a sprawling floor exercise area—establishes an atmosphere of elite athleticism and disciplinary perfection within the vast, professional training hall.", + "seed": 3948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_12/index_6728_seed_3948_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": -0.14004337713122367, + "vlm_logprob": -0.14004337713122367, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.14004337713122367, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04607892036437988 + ], + "opposite_logprob_tokens": [ + -3.42107892036438 + ], + "logprob_reward": -0.04607892036437988, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04607892036437988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049168046563863754, + 0.02613302879035473, + 0.03742501512169838, + 0.03113739937543869, + 0.5905134677886963, + 0.05016358569264412, + 0.13037531077861786, + 0.04453801363706589, + 0.026674935594201088, + 0.013871096074581146 + ], + "credit_map": [ + 0.006645314861088991, + 0.007159282919019461, + 0.006863764952868223, + 0.008403917774558067, + 0.004809590056538582, + 0.0035433487500995398, + 0.00786820612847805, + 0.0038746241480112076, + 0.002822066657245159, + 0.002544972114264965, + 0.0021780836395919323, + 0.003639231901615858, + 0.004391822963953018, + 0.004253513179719448, + 0.0028260962571948767, + 0.0034772432409226894, + 0.00504059623926878, + 0.005961299873888493, + 0.002824563067406416, + 0.004759341944009066, + 0.004112415015697479, + 0.005993140861392021, + 0.0050452095456421375, + 0.003688449738547206, + 0.0027679174672812223, + 0.006297977175563574, + 0.0067694769240915775, + 0.0038768332451581955, + 0.002350836293771863, + 0.003288959851488471, + 0.003147960640490055, + 0.0026374394074082375, + 0.07766376435756683, + 0.20711983740329742, + 0.1442549228668213, + 0.019594987854361534, + 0.023874228820204735, + 0.047295186668634415, + 0.015088343992829323, + 0.05562223121523857, + 0.004395495634526014, + 0.008436167612671852, + 0.005945192184299231, + 0.004708053078502417, + 0.008863665163516998, + 0.005724077578634024, + 0.006245753727853298, + 0.005845179315656424, + 0.017501194030046463, + 0.012430115602910519, + 0.010849075391888618, + 0.023806314915418625, + 0.012544117867946625, + 0.010137234814465046, + 0.022275052964687347, + 0.020832205191254616, + 0.003925084602087736, + 0.0032988605089485645, + 0.009472559206187725, + 0.0038262626621872187, + 0.003930957056581974, + 0.0037845750339329243, + 0.0033593622501939535, + 0.012940351851284504, + 0.0028649151790887117, + 0.00408304063603282, + 0.0015839803963899612, + 0.0025619121734052896, + 0.004962001461535692, + 0.0035376427695155144, + 0.0037565724924206734, + 0.0033248714171350002, + 0.001738616032525897, + 0.0021987860091030598, + 0.0013308924390003085, + 0.001949712517671287, + 0.001738053048029542, + 0.0016451861010864377, + 0.0018327169818803668, + 0.0014371327124536037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.04607892036437988 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211453437805176 + ], + "opposite_logprob_tokens": [ + -1.5021145343780518 + ], + "logprob_reward": -0.25211453437805176, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211453437805176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04436372593045235, + 0.05784191936254501, + 0.02947620302438736, + 0.30300769209861755, + 0.2452690154314041, + 0.0566285140812397, + 0.11316028237342834, + 0.10191795974969864, + 0.022464748471975327, + 0.025869987905025482 + ], + "credit_map": [ + 0.006031407043337822, + 0.006811329163610935, + 0.005351007916033268, + 0.007450580596923828, + 0.0053467582911252975, + 0.0054097301326692104, + 0.005001605488359928, + 0.0029613047372549772, + 0.006003616843372583, + 0.006781524512916803, + 0.004912154283374548, + 0.011705231852829456, + 0.005520893260836601, + 0.009451155550777912, + 0.006260858848690987, + 0.0072064814157783985, + 0.003496602876111865, + 0.0033500450663268566, + 0.003370445454493165, + 0.003111965721473098, + 0.0035797846503555775, + 0.0038963560946285725, + 0.00494358828291297, + 0.003727416042238474, + 0.033280178904533386, + 0.07515834271907806, + 0.07366333901882172, + 0.02069075219333172, + 0.017966939136385918, + 0.029131408780813217, + 0.019504230469465256, + 0.03361246734857559, + 0.033075716346502304, + 0.0896015465259552, + 0.05607212334871292, + 0.008715974166989326, + 0.011264730244874954, + 0.020771542564034462, + 0.007310766726732254, + 0.018456608057022095, + 0.005220906808972359, + 0.005550798494368792, + 0.013708231039345264, + 0.005090837366878986, + 0.005798247642815113, + 0.005846651736646891, + 0.005062953568994999, + 0.010349885560572147, + 0.009767798706889153, + 0.009661886841058731, + 0.01615099608898163, + 0.017075367271900177, + 0.008660521358251572, + 0.010748649016022682, + 0.011820139363408089, + 0.029274923726916313, + 0.014021338894963264, + 0.01069036778062582, + 0.017675181850790977, + 0.010091620497405529, + 0.006446606479585171, + 0.012204146943986416, + 0.005835108458995819, + 0.02495359256863594, + 0.0029836150351911783, + 0.004131601192057133, + 0.00269924639724195, + 0.004430302418768406, + 0.0018030916107818484, + 0.002480318769812584, + 0.0016235498478636146, + 0.0023130232002586126, + 0.002819874556735158, + 0.0046167196705937386, + 0.0055853575468063354, + 0.004341123625636101, + 0.001555047114379704, + 0.003332865657284856, + 0.001528257387690246, + 0.0020907411817461252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.25211453437805176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695101737976074 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.033695101737976074, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695101737976074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05957880988717079, + 0.02786417119204998, + 0.060412585735321045, + 0.2699088752269745, + 0.26830822229385376, + 0.16912783682346344, + 0.057593367993831635, + 0.061490729451179504, + 0.008419398218393326, + 0.017296021804213524 + ], + "credit_map": [ + 0.006286519579589367, + 0.005929574836045504, + 0.009184544906020164, + 0.01063417550176382, + 0.0062845200300216675, + 0.006768130697309971, + 0.008753281086683273, + 0.005738051608204842, + 0.0032130861654877663, + 0.0031048620585352182, + 0.002603613305836916, + 0.0030188830569386482, + 0.003689349628984928, + 0.005015124101191759, + 0.002829058328643441, + 0.004390190355479717, + 0.006548870354890823, + 0.010029952973127365, + 0.0060414201579988, + 0.008504672907292843, + 0.006694798357784748, + 0.00604200130328536, + 0.00903697032481432, + 0.007513890974223614, + 0.030423603951931, + 0.06712567061185837, + 0.06638108193874359, + 0.017085617408156395, + 0.015765085816383362, + 0.026057854294776917, + 0.019654205068945885, + 0.02741573378443718, + 0.03047623112797737, + 0.0890272930264473, + 0.07156582176685333, + 0.010935832746326923, + 0.010709143243730068, + 0.021408505737781525, + 0.006283069495111704, + 0.027902303263545036, + 0.020648006349802017, + 0.04736277088522911, + 0.015865536406636238, + 0.010208770632743835, + 0.016104651615023613, + 0.010978133417665958, + 0.01059580035507679, + 0.037364158779382706, + 0.005476456135511398, + 0.004819116089493036, + 0.008358165621757507, + 0.006937519181519747, + 0.004235449712723494, + 0.005400514230132103, + 0.005458656698465347, + 0.016907483339309692, + 0.008641645312309265, + 0.006276922766119242, + 0.010702691040933132, + 0.006294530350714922, + 0.004162812139838934, + 0.006955296732485294, + 0.003907789476215839, + 0.01454903744161129, + 0.0012213875306770205, + 0.001169999479316175, + 0.0014801480574533343, + 0.0008052982739172876, + 0.0011003867257386446, + 0.0012150787515565753, + 0.0007285694591701031, + 0.0006985298823565245, + 0.0019729649648070335, + 0.002517846878618002, + 0.0018356628715991974, + 0.0035983803682029247, + 0.001582319033332169, + 0.001908017322421074, + 0.0019787882920354605, + 0.00190203997772187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.033695101737976074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134017884731293 + ], + "opposite_logprob_tokens": [ + -1.3134018182754517 + ], + "logprob_reward": -0.3134017884731293, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134017884731293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05875903740525246, + 0.04754042252898216, + 0.032905928790569305, + 0.10261309891939163, + 0.2570376694202423, + 0.16896334290504456, + 0.208538219332695, + 0.07272125035524368, + 0.015129132196307182, + 0.03579183667898178 + ], + "credit_map": [ + 0.009462850168347359, + 0.008330770768225193, + 0.008450670167803764, + 0.009332945570349693, + 0.006012734025716782, + 0.0059702834114432335, + 0.006553585175424814, + 0.004645202308893204, + 0.006179215852171183, + 0.006432367954403162, + 0.00397597486153245, + 0.006448813248425722, + 0.005656661931425333, + 0.010426714085042477, + 0.0033530774526298046, + 0.005067601799964905, + 0.0045037297531962395, + 0.005444658920168877, + 0.003740889485925436, + 0.004081643186509609, + 0.0032662551384419203, + 0.00369702884927392, + 0.004971347749233246, + 0.0032003787346184254, + 0.013712319545447826, + 0.02390911616384983, + 0.023177791386842728, + 0.00585187366232276, + 0.006624927744269371, + 0.01011783629655838, + 0.006172239314764738, + 0.013047005981206894, + 0.038807936012744904, + 0.09885367751121521, + 0.04419827461242676, + 0.01707477495074272, + 0.012927013449370861, + 0.022450316697359085, + 0.009633542969822884, + 0.013092138804495335, + 0.019355248659849167, + 0.04196518659591675, + 0.01636008732020855, + 0.01168348640203476, + 0.019385701045393944, + 0.014970727264881134, + 0.014570902101695538, + 0.030672024935483932, + 0.03188718855381012, + 0.014735420234501362, + 0.024539247155189514, + 0.02726306952536106, + 0.014925820752978325, + 0.02204691618680954, + 0.02593371644616127, + 0.04720686376094818, + 0.011843442916870117, + 0.006607904098927975, + 0.01462997030466795, + 0.005489461589604616, + 0.003895218251273036, + 0.006917242892086506, + 0.003898609196767211, + 0.019439412280917168, + 0.002127290703356266, + 0.002637921366840601, + 0.0021753618493676186, + 0.0019421303877606988, + 0.0015928064240142703, + 0.0018486437620595098, + 0.0013811999233439565, + 0.001423779409378767, + 0.005486905109137297, + 0.00582173140719533, + 0.0033476410899311304, + 0.004483135882765055, + 0.003969945479184389, + 0.006054814904928207, + 0.003251587273553014, + 0.0033760773949325085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.3134017884731293 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054926540702581406 + ], + "opposite_logprob_tokens": [ + -2.929926633834839 + ], + "logprob_reward": -0.054926540702581406, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054926540702581406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10829932987689972, + 0.07279680669307709, + 0.0696139857172966, + 0.1078387051820755, + 0.10804889351129532, + 0.1178804486989975, + 0.1225748136639595, + 0.07862255722284317, + 0.057830389589071274, + 0.1564941555261612 + ], + "credit_map": [ + 0.014619716443121433, + 0.012695273384451866, + 0.014257512986660004, + 0.015399200841784477, + 0.0116981640458107, + 0.013521785847842693, + 0.016514739021658897, + 0.0095929354429245, + 0.007916885428130627, + 0.010297620669007301, + 0.007371982093900442, + 0.011046514846384525, + 0.0076730442233383656, + 0.015029422007501125, + 0.005149096716195345, + 0.008312240242958069, + 0.008419848047196865, + 0.007082931697368622, + 0.006479284726083279, + 0.007927600294351578, + 0.005686514545232058, + 0.012734819203615189, + 0.006181260570883751, + 0.015101732686161995, + 0.015977051109075546, + 0.013939044438302517, + 0.01744309812784195, + 0.011066948994994164, + 0.00849844515323639, + 0.013376954011619091, + 0.010202683508396149, + 0.017334476113319397, + 0.009182509034872055, + 0.010763553902506828, + 0.029514435678720474, + 0.012976537458598614, + 0.00699830986559391, + 0.006737800780683756, + 0.005255203228443861, + 0.026620538905262947, + 0.015630386769771576, + 0.03384549915790558, + 0.011141101829707623, + 0.005541608203202486, + 0.009227200411260128, + 0.01055777259171009, + 0.00727175734937191, + 0.024665117263793945, + 0.016990073025226593, + 0.011084014549851418, + 0.016912268474698067, + 0.018342958763241768, + 0.010137055069208145, + 0.011877622455358505, + 0.018129635602235794, + 0.019101183861494064, + 0.009279481135308743, + 0.009559237398207188, + 0.01244991272687912, + 0.011236533522605896, + 0.0062284283339977264, + 0.01509825699031353, + 0.006442330777645111, + 0.008328376337885857, + 0.0077862185426056385, + 0.008780678734183311, + 0.011301028542220592, + 0.006972160656005144, + 0.005835187155753374, + 0.006592513993382454, + 0.004905835259705782, + 0.005656764376908541, + 0.02059212699532509, + 0.022477908059954643, + 0.01873738318681717, + 0.03454525023698807, + 0.013675018213689327, + 0.01930283196270466, + 0.01173729170113802, + 0.015426339581608772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.054926540702581406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot captures a male athlete's climactic moment, his powerful physique clad in a sleeveless royal blue leotard as he executes a graceful sprint from the gleaming, polished steel parallel bars. With short, black hair damp from exertion and hands coated in white chalk, he thrusts forward decisively, his body spinning in a blur of controlled motion through the air. He lands with pinpoint precision on a thick, navy blue mat, his feet striking the surface simultaneously with a solid, satisfying thud. His knees bend deeply to absorb the force, his arms briefly extended for balance, before he straightens into a resolute, triumphant pose, his expression filled with an intense, focused sense of accomplishment. As he holds his pose, the camera begins a steady, continuous crane shot, tilting upward and pulling back to reveal a vast, well-equipped, modern gymnasium. Bright, diffuse overhead lighting illuminates the space, and the other state-of-the-art equipment gleaming in the background—the pommel horse, rings, and a sprawling floor exercise area—establishes an atmosphere of elite athleticism and disciplinary perfection within the vast, professional training hall.", + "seed": 3949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_12/index_6728_seed_3949_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": -0.07628099136054516, + "vlm_logprob": -0.07628099136054516, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.07628099136054516, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033001579344272614 + ], + "opposite_logprob_tokens": [ + -3.783001661300659 + ], + "logprob_reward": -0.033001579344272614, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.033001579344272614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19807197153568268, + 0.08962046355009079, + 0.0694151446223259, + 0.05983478203415871, + 0.08393170684576035, + 0.1016252189874649, + 0.12407228350639343, + 0.09648919105529785, + 0.143108531832695, + 0.03383074700832367 + ], + "credit_map": [ + 0.023464392870664597, + 0.01787024922668934, + 0.02240569330751896, + 0.023569801822304726, + 0.02236131578683853, + 0.04195394366979599, + 0.02428160421550274, + 0.02216498553752899, + 0.008101261220872402, + 0.011176778934895992, + 0.010955438949167728, + 0.01107101235538721, + 0.009919535368680954, + 0.009893504902720451, + 0.018899668008089066, + 0.009603263810276985, + 0.011756222695112228, + 0.005736358929425478, + 0.010904333554208279, + 0.007657123729586601, + 0.005124724004417658, + 0.007554118521511555, + 0.010217776522040367, + 0.01046448852866888, + 0.007142011076211929, + 0.007138877175748348, + 0.010487418621778488, + 0.005728804040700197, + 0.008187853731215, + 0.006950738839805126, + 0.005374044179916382, + 0.008825032971799374, + 0.00873522087931633, + 0.009263206273317337, + 0.017595568671822548, + 0.007258533034473658, + 0.012458093464374542, + 0.011558223515748978, + 0.008295692503452301, + 0.008767171762883663, + 0.009723799303174019, + 0.012923959642648697, + 0.013420588336884975, + 0.01013017725199461, + 0.016962584108114243, + 0.011948961764574051, + 0.013682008720934391, + 0.012833141721785069, + 0.016224851831793785, + 0.010111678391695023, + 0.012509766966104507, + 0.01895851641893387, + 0.017778806388378143, + 0.012966575101017952, + 0.014818916097283363, + 0.02070317603647709, + 0.01579250581562519, + 0.009256460703909397, + 0.011153721250593662, + 0.009829487651586533, + 0.012172244489192963, + 0.011081702075898647, + 0.015792779624462128, + 0.011410285718739033, + 0.014336999505758286, + 0.01685735583305359, + 0.01706821843981743, + 0.023045595735311508, + 0.022796669974923134, + 0.014089643955230713, + 0.018415279686450958, + 0.01649876870214939, + 0.0031083333306014538, + 0.004083777777850628, + 0.0032778612803667784, + 0.0035151855554431677, + 0.00648635346442461, + 0.004671984817832708, + 0.004636923782527447, + 0.004050326533615589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.033001579344272614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900464534759521 + ], + "opposite_logprob_tokens": [ + -2.5790047645568848 + ], + "logprob_reward": -0.07900464534759521, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900464534759521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22144842147827148, + 0.09895344078540802, + 0.07077176868915558, + 0.08711674064397812, + 0.18571943044662476, + 0.06312615424394608, + 0.07638665288686752, + 0.06240758299827576, + 0.06354070454835892, + 0.07052907347679138 + ], + "credit_map": [ + 0.02311314456164837, + 0.03112405724823475, + 0.036817774176597595, + 0.026995668187737465, + 0.03767523914575577, + 0.023909028619527817, + 0.020961305126547813, + 0.02085219882428646, + 0.007724881172180176, + 0.01398694422096014, + 0.012370518408715725, + 0.014669050462543964, + 0.010273678228259087, + 0.009272956289350986, + 0.02072710543870926, + 0.009928311221301556, + 0.010627303272485733, + 0.007496898528188467, + 0.00834522396326065, + 0.0076690721325576305, + 0.006274085957556963, + 0.010648355819284916, + 0.013016531243920326, + 0.006694299168884754, + 0.009527083486318588, + 0.01079577300697565, + 0.01393144577741623, + 0.01331594493240118, + 0.010406018234789371, + 0.00789306778460741, + 0.009680653922259808, + 0.011566750705242157, + 0.01470356248319149, + 0.02095971256494522, + 0.04786647856235504, + 0.016767805442214012, + 0.016225574538111687, + 0.01617009937763214, + 0.019792350009083748, + 0.033233851194381714, + 0.006110723130404949, + 0.0073828017339110374, + 0.007544129155576229, + 0.007501588203012943, + 0.008621320128440857, + 0.007702034432440996, + 0.010039200074970722, + 0.008224361576139927, + 0.011294364929199219, + 0.00669245095923543, + 0.010777796618640423, + 0.011506859213113785, + 0.00799145083874464, + 0.008282710798084736, + 0.008993777446448803, + 0.010847232304513454, + 0.011198557913303375, + 0.005307689309120178, + 0.009832966141402721, + 0.006986342836171389, + 0.005784917622804642, + 0.008402900770306587, + 0.008719092234969139, + 0.0061751133762300014, + 0.006992862094193697, + 0.006601438391953707, + 0.006241569761186838, + 0.007769505959004164, + 0.010338720865547657, + 0.006465771701186895, + 0.009949682280421257, + 0.009181150235235691, + 0.007213226519525051, + 0.008237091824412346, + 0.009054972790181637, + 0.005510299000889063, + 0.015009630471467972, + 0.010527616366744041, + 0.006476681679487228, + 0.00849955528974533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07900464534759521 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038091905415058136 + ], + "opposite_logprob_tokens": [ + -3.2880918979644775 + ], + "logprob_reward": -0.038091905415058136, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038091905415058136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16815465688705444, + 0.1383548229932785, + 0.08919946104288101, + 0.0954655334353447, + 0.1872636377811432, + 0.07516460120677948, + 0.11475040018558502, + 0.0320846289396286, + 0.05784058943390846, + 0.04172160476446152 + ], + "credit_map": [ + 0.022992892190814018, + 0.02286943607032299, + 0.01601676270365715, + 0.016381174325942993, + 0.023410970345139503, + 0.030753986909985542, + 0.016104059293866158, + 0.019625378772616386, + 0.011504332534968853, + 0.020694326609373093, + 0.016658371314406395, + 0.021804170683026314, + 0.012918503023684025, + 0.013900067657232285, + 0.0275440514087677, + 0.013331014662981033, + 0.010623831301927567, + 0.00808213371783495, + 0.013013966381549835, + 0.011254212819039822, + 0.012322784401476383, + 0.009222741238772869, + 0.016502121463418007, + 0.00817768182605505, + 0.012587719596922398, + 0.00886853039264679, + 0.016230465844273567, + 0.010604056529700756, + 0.009269602596759796, + 0.01580306701362133, + 0.009621492587029934, + 0.012480608187615871, + 0.01459942851215601, + 0.016401760280132294, + 0.051588863134384155, + 0.017204126343131065, + 0.012836867943406105, + 0.019955987110733986, + 0.014514892362058163, + 0.04016171395778656, + 0.004448542837053537, + 0.010545127093791962, + 0.007085725665092468, + 0.008740035817027092, + 0.01682964712381363, + 0.007177548483014107, + 0.01138208620250225, + 0.008955893106758595, + 0.010690453462302685, + 0.008514760993421078, + 0.0179985873401165, + 0.018773160874843597, + 0.00926959328353405, + 0.020002424716949463, + 0.014340479858219624, + 0.015160957351326942, + 0.004474584013223648, + 0.003022601595148444, + 0.005029027350246906, + 0.004619809798896313, + 0.003504154970869422, + 0.0033459924161434174, + 0.004613546654582024, + 0.0034749156329780817, + 0.00905666034668684, + 0.005807996727526188, + 0.007305442355573177, + 0.009578476659953594, + 0.006198284216225147, + 0.005445081740617752, + 0.005703323055058718, + 0.008745328523218632, + 0.0037585662212222815, + 0.0047370800748467445, + 0.004573441110551357, + 0.004948039539158344, + 0.008571777492761612, + 0.005931471008807421, + 0.00499268714338541, + 0.00420854426920414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.038091905415058136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015266716480255 + ], + "opposite_logprob_tokens": [ + -1.7015266418457031 + ], + "logprob_reward": -0.2015266716480255, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015266716480255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26335862278938293, + 0.08266724646091461, + 0.12364619970321655, + 0.08871462941169739, + 0.11361966282129288, + 0.06274546682834625, + 0.06520172208547592, + 0.05642041936516762, + 0.052518028765916824, + 0.09110800921916962 + ], + "credit_map": [ + 0.03342483192682266, + 0.03020476922392845, + 0.03848978132009506, + 0.028167128562927246, + 0.03058135136961937, + 0.04726285859942436, + 0.03000124916434288, + 0.025226637721061707, + 0.008744319900870323, + 0.014611390419304371, + 0.011954705230891705, + 0.009862122125923634, + 0.01082286611199379, + 0.007569499313831329, + 0.011995483189821243, + 0.007106861099600792, + 0.020638592541217804, + 0.010107448324561119, + 0.021581001579761505, + 0.013790104538202286, + 0.012724815867841244, + 0.014593844301998615, + 0.013909528031945229, + 0.016300860792398453, + 0.008096273988485336, + 0.006556807551532984, + 0.014085349626839161, + 0.01634141430258751, + 0.009565106593072414, + 0.008156530559062958, + 0.014721754938364029, + 0.011191385798156261, + 0.0075242710299789906, + 0.017060207203030586, + 0.021738126873970032, + 0.015463318675756454, + 0.017465578392148018, + 0.008371720090508461, + 0.010488525964319706, + 0.015507909469306469, + 0.0073046390898525715, + 0.006201732903718948, + 0.007729274686425924, + 0.006859824992716312, + 0.009988025762140751, + 0.008005882613360882, + 0.009302563965320587, + 0.007353522349148989, + 0.005160408094525337, + 0.004732571542263031, + 0.011498035863041878, + 0.007159546483308077, + 0.005227229557931423, + 0.01481922622770071, + 0.008895056322216988, + 0.007709643337875605, + 0.007294275797903538, + 0.00633644312620163, + 0.007193543948233128, + 0.007048404775559902, + 0.00565467169508338, + 0.006311612669378519, + 0.008379591628909111, + 0.008201876655220985, + 0.006764738820493221, + 0.006004187744110823, + 0.005133010447025299, + 0.009139657951891422, + 0.0046349563635885715, + 0.007958769798278809, + 0.004585232585668564, + 0.008297477848827839, + 0.0087119797244668, + 0.010248912498354912, + 0.011311636306345463, + 0.008073108270764351, + 0.019970620051026344, + 0.009917842224240303, + 0.011751776561141014, + 0.011122130788862705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.2015266716480255 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a male athlete wearing a sleeveless royal blue leotard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the athlete perform a sprint and spin off gleaming, polished steel parallel bars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the athlete land on a thick, navy blue mat with both feet striking simultaneously?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a steady crane shot that tilts upward and pulls back to reveal a modern gymnasium?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible gymnasium equipped with a pommel horse, rings, and a floor exercise area?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780155047774315 + ], + "opposite_logprob_tokens": [ + -3.529780149459839 + ], + "logprob_reward": -0.029780155047774315, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780155047774315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33748942613601685, + 0.06937805563211441, + 0.09595471620559692, + 0.08879289031028748, + 0.15020650625228882, + 0.049818120896816254, + 0.06503023952245712, + 0.0667712390422821, + 0.046012867242097855, + 0.03054598905146122 + ], + "credit_map": [ + 0.043240081518888474, + 0.03250398486852646, + 0.030077863484621048, + 0.027621882036328316, + 0.04274122789502144, + 0.08424399048089981, + 0.04049539938569069, + 0.036565035581588745, + 0.0054726204834878445, + 0.009266236796975136, + 0.008274977095425129, + 0.011201000772416592, + 0.006900765933096409, + 0.007066958583891392, + 0.014424101449549198, + 0.006771396845579147, + 0.01921124756336212, + 0.009165806695818901, + 0.017758740112185478, + 0.010143245570361614, + 0.006440155208110809, + 0.013034735806286335, + 0.007367877289652824, + 0.012832904234528542, + 0.01269717700779438, + 0.009127638302743435, + 0.021458463743329048, + 0.010763384401798248, + 0.006540756672620773, + 0.00923379510641098, + 0.006878699641674757, + 0.012092973105609417, + 0.024990472942590714, + 0.011612586677074432, + 0.030454909428954124, + 0.01208209153264761, + 0.009483831003308296, + 0.02897094376385212, + 0.013354157097637653, + 0.01925751194357872, + 0.008665493689477444, + 0.004301445558667183, + 0.006573319435119629, + 0.0043479073792696, + 0.00375283882021904, + 0.007813497446477413, + 0.004431969486176968, + 0.009931649081408978, + 0.008974866941571236, + 0.0056051164865493774, + 0.011105729267001152, + 0.00692357774823904, + 0.0043545677326619625, + 0.01553231943398714, + 0.007279231213033199, + 0.005254833027720451, + 0.013230456970632076, + 0.009270201437175274, + 0.007077978458255529, + 0.006849053781479597, + 0.006153458263725042, + 0.011238831095397472, + 0.004996121861040592, + 0.00795513391494751, + 0.0037263117264956236, + 0.004416218958795071, + 0.005746804643422365, + 0.007312101777642965, + 0.008655251935124397, + 0.004808626603335142, + 0.006776823662221432, + 0.004570727702230215, + 0.004380953498184681, + 0.0036163688637316227, + 0.004734245594590902, + 0.0025873284321278334, + 0.0040243035182356834, + 0.003916092216968536, + 0.0027955006808042526, + 0.004491196479648352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029780155047774315 + } + ], + "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/0000110/group_13/index_tensor([7512, 7512])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_tensor([7512, 7512])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7dad0a5ee644b43b096580cc2de5061975d6f3fa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_tensor([7512, 7512])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid, uninterrupted tracking shot, maintaining an eye-level perspective with its subject, follows a paper airplane as it gracefully navigates a vast, desolate modern office. Meticulously folded from a single sheet of crisp white paper, its sharp creases forming the classic silhouette of a boomerang, the airplane glides silently and effortlessly through the still air. Cool, sterile light from long fluorescent lamps embedded in the ceiling bathes the scene, casting soft, elongated shadows and creating sharp, mirrored highlights on the polished concrete floors and smooth tabletops. The camera soars along a serene trajectory over rows of empty, minimalist desks, flying past the dark, reflective surfaces of powered-off computer screens and neatly arranged keyboards. Its gentle tilt, its uninterrupted motion contrasting with the static surroundings of abandoned ergonomic chairs, creates an atmosphere of profound tranquility and quiet solitude, a moment of simple, poetic beauty capturing the quiet heart of the corporate world.", + "seed": 14826, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_7512_seed_14826_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": -0.22054447717964648, + "vlm_logprob": -0.22054447717964648, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.22054447717964648, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5793482065200806 + ], + "opposite_logprob_tokens": [ + -0.8293482065200806 + ], + "logprob_reward": -0.5793482065200806, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5793482065200806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18617621064186096, + 0.17364506423473358, + 0.039722468703985214, + 0.0432429239153862, + 0.09828808158636093, + 0.04877980425953865, + 0.04628276824951172, + 0.10775478929281235, + 0.05083636939525604, + 0.20527148246765137 + ], + "credit_map": [ + 0.016073863953351974, + 0.01130713615566492, + 0.014075418002903461, + 0.041063517332077026, + 0.009857297874987125, + 0.018222972750663757, + 0.057964786887168884, + 0.01761120744049549, + 0.017932545393705368, + 0.015563510358333588, + 0.03138108178973198, + 0.028752697631716728, + 0.016115086153149605, + 0.018138471990823746, + 0.022591037675738335, + 0.02317063882946968, + 0.00643262080848217, + 0.0038378587923943996, + 0.00399426743388176, + 0.005053685978055, + 0.003021923126652837, + 0.005393038038164377, + 0.005078955087810755, + 0.006910115480422974, + 0.009223983623087406, + 0.0062111238949000835, + 0.0031371889635920525, + 0.005078413523733616, + 0.005573011003434658, + 0.004225508309900761, + 0.005615269299596548, + 0.004178423434495926, + 0.019097905606031418, + 0.007876654155552387, + 0.01042726170271635, + 0.01878400705754757, + 0.005522801540791988, + 0.013413081876933575, + 0.01252498384565115, + 0.010641377419233322, + 0.0043321410194039345, + 0.0046123843640089035, + 0.006674592383205891, + 0.006967855151742697, + 0.00705719506368041, + 0.007247146684676409, + 0.007149884477257729, + 0.0047386023215949535, + 0.006053875666111708, + 0.00413571298122406, + 0.00513125117868185, + 0.004903165623545647, + 0.004396277479827404, + 0.00481270020827651, + 0.00826776959002018, + 0.008582020178437233, + 0.011933069676160812, + 0.02692163735628128, + 0.010047622956335545, + 0.013093573041260242, + 0.008711672388017178, + 0.013649213127791882, + 0.012779501266777515, + 0.010618507862091064, + 0.007102106232196093, + 0.006515003275126219, + 0.005349113140255213, + 0.006760955788195133, + 0.005984147544950247, + 0.005002300720661879, + 0.007264873944222927, + 0.006857871077954769, + 0.0077432794496417046, + 0.009917139075696468, + 0.01837742142379284, + 0.08423766493797302, + 0.014375263825058937, + 0.0469169057905674, + 0.011173243634402752, + 0.01253055315464735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.5793482065200806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871406018733978 + ], + "opposite_logprob_tokens": [ + -1.1371406316757202 + ], + "logprob_reward": -0.3871406018733978, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871406018733978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14497745037078857, + 0.1903219074010849, + 0.04317512735724449, + 0.03877847269177437, + 0.06971249729394913, + 0.01839497685432434, + 0.08452322334051132, + 0.07145541161298752, + 0.036020662635564804, + 0.30264028906822205 + ], + "credit_map": [ + 0.01740086078643799, + 0.009767012670636177, + 0.010798565112054348, + 0.031199438497424126, + 0.010483259335160255, + 0.01226870622485876, + 0.04042134806513786, + 0.012638263404369354, + 0.023448491469025612, + 0.01638101600110531, + 0.029170716181397438, + 0.029553450644016266, + 0.01482338085770607, + 0.027527233585715294, + 0.027993395924568176, + 0.02142423577606678, + 0.00605968851596117, + 0.006612344644963741, + 0.003331931075081229, + 0.005765791982412338, + 0.004413127433508635, + 0.005302858538925648, + 0.005276308860629797, + 0.006413082592189312, + 0.009286305867135525, + 0.006101501174271107, + 0.003181210719048977, + 0.0033681110944598913, + 0.003378430614247918, + 0.00491178035736084, + 0.0044467574916779995, + 0.0041043758392333984, + 0.012065134942531586, + 0.0065083131194114685, + 0.008923596702516079, + 0.009846086613833904, + 0.00989415030926466, + 0.008058778941631317, + 0.009154058061540127, + 0.005262386053800583, + 0.0016136136837303638, + 0.0020440248772501945, + 0.0030253923032432795, + 0.0025030418764799833, + 0.002702789381146431, + 0.0023387656547129154, + 0.0023162467405200005, + 0.0018511043163016438, + 0.00684422766789794, + 0.008147578686475754, + 0.006006724666804075, + 0.008513031527400017, + 0.007758057676255703, + 0.005769787356257439, + 0.026511531323194504, + 0.014972292818129063, + 0.00753041822463274, + 0.019068151712417603, + 0.008404923602938652, + 0.00858917087316513, + 0.005749408155679703, + 0.008004707284271717, + 0.0077042351476848125, + 0.006404402665793896, + 0.0036638081073760986, + 0.0038044112734496593, + 0.004840161185711622, + 0.008157174102962017, + 0.003316627349704504, + 0.003267170861363411, + 0.005490322131663561, + 0.0034809920471161604, + 0.008019771426916122, + 0.010863217525184155, + 0.027532989159226418, + 0.13192392885684967, + 0.011639582924544811, + 0.0813271552324295, + 0.017198270186781883, + 0.014135410077869892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.3871406018733978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789751261472702 + ], + "opposite_logprob_tokens": [ + -2.578975200653076 + ], + "logprob_reward": -0.0789751261472702, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789751261472702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08216594159603119, + 0.1011582612991333, + 0.1244456097483635, + 0.047440823167562485, + 0.0478297658264637, + 0.028970198705792427, + 0.039825815707445145, + 0.12816931307315826, + 0.02345305308699608, + 0.37654128670692444 + ], + "credit_map": [ + 0.009846003726124763, + 0.00986574124544859, + 0.010679468512535095, + 0.00982746947556734, + 0.008243494667112827, + 0.012074381113052368, + 0.012702125124633312, + 0.008927248418331146, + 0.020651284605264664, + 0.008950785733759403, + 0.02087220922112465, + 0.008592560887336731, + 0.005678139626979828, + 0.009925685822963715, + 0.013711719773709774, + 0.012775861658155918, + 0.021922489628195763, + 0.017392676323652267, + 0.009997906163334846, + 0.01616509072482586, + 0.013558963313698769, + 0.018211854621767998, + 0.014039365574717522, + 0.013157252222299576, + 0.006340859457850456, + 0.005239987280219793, + 0.00333611061796546, + 0.006119781639426947, + 0.004167567472904921, + 0.008722004480659962, + 0.008240416646003723, + 0.005274089053273201, + 0.009344528429210186, + 0.004077303688973188, + 0.006553844548761845, + 0.007603958249092102, + 0.005400441121309996, + 0.004987876396626234, + 0.005478890612721443, + 0.00438291672617197, + 0.002888544462621212, + 0.0034040140453726053, + 0.0049510495737195015, + 0.004169634077697992, + 0.003934390842914581, + 0.003847946645691991, + 0.0034952859859913588, + 0.0022793319076299667, + 0.005007529631257057, + 0.0035518871154636145, + 0.004474148154258728, + 0.005746547132730484, + 0.004334262106567621, + 0.004683313891291618, + 0.005835444666445255, + 0.006192678585648537, + 0.014999963343143463, + 0.01198474783450365, + 0.016918769106268883, + 0.021365340799093246, + 0.009584683924913406, + 0.01905973255634308, + 0.02056456170976162, + 0.013691496104001999, + 0.006085679866373539, + 0.0025079064071178436, + 0.001989500829949975, + 0.002478920156136155, + 0.00282462639734149, + 0.0028003521729260683, + 0.0027066755574196577, + 0.0020593886729329824, + 0.011119320057332516, + 0.014281956478953362, + 0.03545619174838066, + 0.16445302963256836, + 0.015303974971175194, + 0.09674286842346191, + 0.020373933017253876, + 0.01880996860563755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.0789751261472702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10165145993232727, + 0.08219601213932037, + 0.061490342020988464, + 0.06511958688497543, + 0.06396747380495071, + 0.03533969819545746, + 0.05014977976679802, + 0.4140560030937195, + 0.026679523289203644, + 0.09935016185045242 + ], + "credit_map": [ + 0.011549717746675014, + 0.011730865575373173, + 0.012787209823727608, + 0.013011611998081207, + 0.010507624596357346, + 0.014157441444694996, + 0.017203276976943016, + 0.010703705251216888, + 0.014364163391292095, + 0.005841781850904226, + 0.014365284703671932, + 0.009142962284386158, + 0.0047827837988734245, + 0.011921179480850697, + 0.010698841884732246, + 0.011079005897045135, + 0.01007105503231287, + 0.0066810185089707375, + 0.0042031495831906796, + 0.007927956990897655, + 0.007955511100590229, + 0.009380295872688293, + 0.008281759917736053, + 0.006989590357989073, + 0.00958098005503416, + 0.006198234856128693, + 0.00452010752633214, + 0.010081261396408081, + 0.007362679112702608, + 0.00924625713378191, + 0.012355836108326912, + 0.005774228367954493, + 0.012367590330541134, + 0.005868321284651756, + 0.006817139219492674, + 0.00865270011126995, + 0.005715060047805309, + 0.007120165042579174, + 0.008514182642102242, + 0.008912314660847187, + 0.0025959473568946123, + 0.003426177427172661, + 0.005111899692565203, + 0.005154329352080822, + 0.004058567341417074, + 0.007350839674472809, + 0.004936488810926676, + 0.0027054448146373034, + 0.004559542052447796, + 0.005965034477412701, + 0.005105046555399895, + 0.006801018491387367, + 0.005772932432591915, + 0.0047230347990989685, + 0.009668446145951748, + 0.0075547220185399055, + 0.04207340627908707, + 0.05662733316421509, + 0.05424346774816513, + 0.05317451432347298, + 0.03063221648335457, + 0.0774088203907013, + 0.056875504553318024, + 0.04302072897553444, + 0.004658600315451622, + 0.003873657202348113, + 0.0027140937745571136, + 0.0028790913056582212, + 0.003216415410861373, + 0.0027173387352377176, + 0.002970220288261771, + 0.0036501057911664248, + 0.0043732221238315105, + 0.005818482954055071, + 0.009120950475335121, + 0.039193045347929, + 0.00833665207028389, + 0.020827898755669594, + 0.0057983435690402985, + 0.005881558172404766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.014206582680344582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305186867713928 + ], + "opposite_logprob_tokens": [ + -3.1680519580841064 + ], + "logprob_reward": -0.04305186867713928, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305186867713928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08591018617153168, + 0.06801007688045502, + 0.0685584619641304, + 0.059661779552698135, + 0.03797417879104614, + 0.02847111225128174, + 0.04072342813014984, + 0.16392190754413605, + 0.0351729616522789, + 0.4115959405899048 + ], + "credit_map": [ + 0.012943259440362453, + 0.008394721895456314, + 0.009364747442305088, + 0.006689735222607851, + 0.010740406811237335, + 0.01499837078154087, + 0.01270029041916132, + 0.010078654624521732, + 0.008689392358064651, + 0.00779323186725378, + 0.00732678035274148, + 0.008617290295660496, + 0.00454225717112422, + 0.008947283960878849, + 0.014057077467441559, + 0.008036765269935131, + 0.012443791143596172, + 0.010166014544665813, + 0.006038590334355831, + 0.009028635919094086, + 0.006011333782225847, + 0.00874229148030281, + 0.007911883294582367, + 0.008215916343033314, + 0.016172870993614197, + 0.0116250179708004, + 0.003263256046921015, + 0.00568943377584219, + 0.004163564182817936, + 0.0062911417335271835, + 0.005903067532926798, + 0.00655342685058713, + 0.005755052901804447, + 0.003507337998598814, + 0.004683343227952719, + 0.004632849711924791, + 0.005454741418361664, + 0.004277374595403671, + 0.005340052768588066, + 0.004323423374444246, + 0.003024305682629347, + 0.002505356213077903, + 0.004482871852815151, + 0.0042035444639623165, + 0.003618652233853936, + 0.005368654616177082, + 0.0033272532746195793, + 0.0019404739141464233, + 0.005179658532142639, + 0.0024018881376832724, + 0.004117053002119064, + 0.004218980669975281, + 0.0037004246842116117, + 0.00413210503757, + 0.007518046535551548, + 0.009455271996557713, + 0.01511323917657137, + 0.020084718242287636, + 0.020784523338079453, + 0.02400241792201996, + 0.01252949982881546, + 0.02903113141655922, + 0.025216100737452507, + 0.01716027408838272, + 0.006483462639153004, + 0.0031816475093364716, + 0.0029559568502008915, + 0.005460395943373442, + 0.004451748915016651, + 0.0038689358625561, + 0.004677911289036274, + 0.00409290287643671, + 0.011445286683738232, + 0.01577533781528473, + 0.039099231362342834, + 0.17915143072605133, + 0.01865808665752411, + 0.10451958328485489, + 0.02231431007385254, + 0.020632660016417503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.04305186867713928 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, uninterrupted tracking shot, maintaining an eye-level perspective with its subject, follows a paper airplane as it gracefully navigates a vast, desolate modern office. Meticulously folded from a single sheet of crisp white paper, its sharp creases forming the classic silhouette of a boomerang, the airplane glides silently and effortlessly through the still air. Cool, sterile light from long fluorescent lamps embedded in the ceiling bathes the scene, casting soft, elongated shadows and creating sharp, mirrored highlights on the polished concrete floors and smooth tabletops. The camera soars along a serene trajectory over rows of empty, minimalist desks, flying past the dark, reflective surfaces of powered-off computer screens and neatly arranged keyboards. Its gentle tilt, its uninterrupted motion contrasting with the static surroundings of abandoned ergonomic chairs, creates an atmosphere of profound tranquility and quiet solitude, a moment of simple, poetic beauty capturing the quiet heart of the corporate world.", + "seed": 14827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_7512_seed_14827_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": -0.055160497315227985, + "vlm_logprob": -0.055160497315227985, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.055160497315227985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051190268248319626 + ], + "opposite_logprob_tokens": [ + -3.0511903762817383 + ], + "logprob_reward": -0.051190268248319626, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051190268248319626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29213032126426697, + 0.023130035027861595, + 0.17743223905563354, + 0.09086941927671432, + 0.04475933313369751, + 0.06075382977724075, + 0.024545174092054367, + 0.05391818657517433, + 0.06419361382722855, + 0.16826783120632172 + ], + "credit_map": [ + 0.015064755454659462, + 0.016482418403029442, + 0.0181418526917696, + 0.04085444286465645, + 0.017868734896183014, + 0.01550590805709362, + 0.09302762150764465, + 0.07518459111452103, + 0.0011701456969603896, + 0.0023463477846235037, + 0.004136787727475166, + 0.004203748423606157, + 0.0015492151724174619, + 0.0020842193625867367, + 0.004651293158531189, + 0.0029882784001529217, + 0.006477142218500376, + 0.005964695475995541, + 0.007635877467691898, + 0.06705086678266525, + 0.0031327896285802126, + 0.007457244209945202, + 0.012950187548995018, + 0.06676344573497772, + 0.0051422202959656715, + 0.009097426198422909, + 0.010471629910171032, + 0.014309006743133068, + 0.01468904409557581, + 0.008728872053325176, + 0.015407450497150421, + 0.013023768551647663, + 0.0047200522385537624, + 0.004676355980336666, + 0.002982634585350752, + 0.004088090732693672, + 0.0027463436126708984, + 0.004943835083395243, + 0.009753823280334473, + 0.01084819994866848, + 0.0057267616502940655, + 0.005174538120627403, + 0.015320397913455963, + 0.0048021962866187096, + 0.003807992208749056, + 0.0070694731548428535, + 0.005151932593435049, + 0.013700535520911217, + 0.002866489579901099, + 0.0020553963258862495, + 0.0048477244563400745, + 0.0029370335396379232, + 0.0018815883668139577, + 0.003226945176720619, + 0.0039206198416650295, + 0.002809376223012805, + 0.003980264998972416, + 0.0035181057173758745, + 0.0033624754287302494, + 0.005024631042033434, + 0.008422911167144775, + 0.011712200939655304, + 0.00920850783586502, + 0.0086890934035182, + 0.007610394153743982, + 0.008071670308709145, + 0.007695783860981464, + 0.00800191517919302, + 0.005729565862566233, + 0.01051557157188654, + 0.00624652486294508, + 0.010322190821170807, + 0.030490590259432793, + 0.013845321722328663, + 0.018180496990680695, + 0.020283756777644157, + 0.02674170397222042, + 0.015509559772908688, + 0.022590726613998413, + 0.02062567137181759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.051190268248319626 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038385033607483 + ], + "opposite_logprob_tokens": [ + -1.9103838205337524 + ], + "logprob_reward": -0.16038385033607483, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038385033607483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21225854754447937, + 0.022913824766874313, + 0.05355381965637207, + 0.0589209645986557, + 0.07357750087976456, + 0.04471738263964653, + 0.04682276397943497, + 0.06609464436769485, + 0.2768574059009552, + 0.14428317546844482 + ], + "credit_map": [ + 0.01677180454134941, + 0.020250730216503143, + 0.02776961959898472, + 0.03249608352780342, + 0.017697131261229515, + 0.015254724770784378, + 0.04990367963910103, + 0.032114773988723755, + 0.0018617482855916023, + 0.002805229974910617, + 0.0021220699418336153, + 0.004182552918791771, + 0.0016616210341453552, + 0.0018883076263591647, + 0.00507377227768302, + 0.003318521659821272, + 0.002249796176329255, + 0.002137879841029644, + 0.0034516530577093363, + 0.015849197283387184, + 0.0019529052078723907, + 0.005355208180844784, + 0.0060448599979281425, + 0.016512319445610046, + 0.005216598976403475, + 0.0064305528067052364, + 0.007824084721505642, + 0.00890039186924696, + 0.008644518442451954, + 0.005229436792433262, + 0.00874989666044712, + 0.007925484329462051, + 0.011913776397705078, + 0.008954203687608242, + 0.005428063217550516, + 0.00557553069666028, + 0.007040564902126789, + 0.008598425425589085, + 0.010853316634893417, + 0.015213621780276299, + 0.004561669658869505, + 0.0037496474105864763, + 0.01034604012966156, + 0.0036776370834559202, + 0.0029795991722494364, + 0.0056522260420024395, + 0.0036651312839239836, + 0.010085432790219784, + 0.00470726890489459, + 0.0038306459318846464, + 0.007893801666796207, + 0.006115282885730267, + 0.004498462192714214, + 0.006770233623683453, + 0.007126379292458296, + 0.0058806887827813625, + 0.004424015060067177, + 0.005341587588191032, + 0.004946405068039894, + 0.005937905516475439, + 0.013931693509221077, + 0.009384263306856155, + 0.01165513414889574, + 0.010473634116351604, + 0.027543891221284866, + 0.022507498040795326, + 0.02126035839319229, + 0.04133852943778038, + 0.030774829909205437, + 0.04685947299003601, + 0.04478829726576805, + 0.041784532368183136, + 0.02324260026216507, + 0.015562147833406925, + 0.016705051064491272, + 0.018578991293907166, + 0.020148474723100662, + 0.013355555944144726, + 0.020333703607320786, + 0.01635664515197277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.16038385033607483 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.45923900604248047, + 0.09634190797805786, + 0.13398770987987518, + 0.021730074658989906, + 0.03819732367992401, + 0.05044887214899063, + 0.025405840948224068, + 0.05132337659597397, + 0.04175541549921036, + 0.08157048374414444 + ], + "credit_map": [ + 0.02495107240974903, + 0.016239799559116364, + 0.02616449072957039, + 0.05978861823678017, + 0.022808071225881577, + 0.020410588011145592, + 0.15386171638965607, + 0.1350146234035492, + 0.005364646669477224, + 0.007731468882411718, + 0.021171215921640396, + 0.01750492863357067, + 0.0068042464554309845, + 0.008565720170736313, + 0.015895115211606026, + 0.013304568827152252, + 0.0049961828626692295, + 0.004478435032069683, + 0.005578169133514166, + 0.04951665922999382, + 0.0029246979393064976, + 0.0058566308580338955, + 0.010565860196948051, + 0.050071071833372116, + 0.0028773306403309107, + 0.002473828149959445, + 0.003275437979027629, + 0.0020070711616426706, + 0.0027751512825489044, + 0.001922914874739945, + 0.003398067783564329, + 0.0030002733692526817, + 0.006673174910247326, + 0.004899566061794758, + 0.003674349281936884, + 0.004083716776221991, + 0.0030955588445067406, + 0.004665843676775694, + 0.005174637772142887, + 0.005930477753281593, + 0.0048553310334682465, + 0.003926449920982122, + 0.014098795130848885, + 0.00460407417267561, + 0.0027499787975102663, + 0.005650815553963184, + 0.0036081587895751, + 0.010955270379781723, + 0.002753789536654949, + 0.002119791926816106, + 0.005585722159594297, + 0.0026822867803275585, + 0.0017877042992040515, + 0.0030595671851187944, + 0.003814286319538951, + 0.0036026937887072563, + 0.004573943093419075, + 0.00491683091968298, + 0.003876808565109968, + 0.005424181930720806, + 0.006525016389787197, + 0.010167467407882214, + 0.007676645647734404, + 0.00816248171031475, + 0.004352156538516283, + 0.004390842281281948, + 0.003232779912650585, + 0.005619965493679047, + 0.004337542690336704, + 0.007532985880970955, + 0.006122476886957884, + 0.006166664883494377, + 0.012313039042055607, + 0.006074955686926842, + 0.010267147794365883, + 0.010384595021605492, + 0.010997897014021873, + 0.00908657070249319, + 0.011625297367572784, + 0.01082097552716732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012537825852632523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646667934954166 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.008646667934954166, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646667934954166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42197296023368835, + 0.08264899253845215, + 0.07219098508358002, + 0.03364456817507744, + 0.031981829553842545, + 0.025886666029691696, + 0.045996878296136856, + 0.0922236368060112, + 0.09980057924985886, + 0.0936528667807579 + ], + "credit_map": [ + 0.031220853328704834, + 0.021577496081590652, + 0.0257803313434124, + 0.07052890211343765, + 0.027046436443924904, + 0.01822986826300621, + 0.1406552940607071, + 0.08693379908800125, + 0.00566968834027648, + 0.00697690574452281, + 0.009448947384953499, + 0.012867347337305546, + 0.008019158616662025, + 0.00873174611479044, + 0.016454126685857773, + 0.014481066726148129, + 0.0029906430281698704, + 0.0027067784685641527, + 0.0036963196471333504, + 0.025963664054870605, + 0.0021001878194510937, + 0.0034449156373739243, + 0.005574352573603392, + 0.025714127346873283, + 0.0018173801945522428, + 0.0029567244928330183, + 0.0044983453117311, + 0.005284131970256567, + 0.0050732907839119434, + 0.003230396658182144, + 0.005715289618819952, + 0.005069008097052574, + 0.0046761552803218365, + 0.003883433062583208, + 0.0027431028429418802, + 0.004166324622929096, + 0.0027204006910324097, + 0.002892846241593361, + 0.004350788425654173, + 0.006548777222633362, + 0.0025864357594400644, + 0.002551543526351452, + 0.006578573491424322, + 0.003271087072789669, + 0.0014090238837525249, + 0.002710356842726469, + 0.0024603954516351223, + 0.004319249652326107, + 0.005536187440156937, + 0.0029239889699965715, + 0.009050155989825726, + 0.006522784009575844, + 0.0037653292529284954, + 0.006058107130229473, + 0.0068589732982218266, + 0.005281352903693914, + 0.008072045631706715, + 0.007915248163044453, + 0.006927101407200098, + 0.010978858917951584, + 0.008336449041962624, + 0.023741472512483597, + 0.011420362628996372, + 0.014832094311714172, + 0.01044376753270626, + 0.01174208428710699, + 0.007132903207093477, + 0.01395796611905098, + 0.01063071470707655, + 0.017661627382040024, + 0.013602256774902344, + 0.014629263430833817, + 0.014392835088074207, + 0.0075742751359939575, + 0.01202209573239088, + 0.013569734059274197, + 0.011889269575476646, + 0.010463256388902664, + 0.01210959441959858, + 0.01163181196898222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.008646667934954166 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304387420415878 + ], + "opposite_logprob_tokens": [ + -3.168043851852417 + ], + "logprob_reward": -0.04304387420415878, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304387420415878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48795372247695923, + 0.061242666095495224, + 0.12579090893268585, + 0.03561525419354439, + 0.04241017997264862, + 0.035305753350257874, + 0.04408218339085579, + 0.0373309925198555, + 0.05969012528657913, + 0.07057823985815048 + ], + "credit_map": [ + 0.032114192843437195, + 0.048620786517858505, + 0.04983648657798767, + 0.058860089629888535, + 0.03595250844955444, + 0.030012767761945724, + 0.12650763988494873, + 0.10604923218488693, + 0.004186860751360655, + 0.005436290055513382, + 0.007229618262499571, + 0.009442820213735104, + 0.004901854787021875, + 0.005857941694557667, + 0.013389545492827892, + 0.010797737166285515, + 0.004550919868052006, + 0.004641147330403328, + 0.004973483271896839, + 0.04681452736258507, + 0.0028537793550640345, + 0.005400901660323143, + 0.009142067283391953, + 0.04741407930850983, + 0.004124338272958994, + 0.004889633506536484, + 0.004810719285160303, + 0.004366715904325247, + 0.004983855411410332, + 0.002976805903017521, + 0.004553254693746567, + 0.0049099307507276535, + 0.005823607090860605, + 0.005966430064290762, + 0.002919253893196583, + 0.005525344051420689, + 0.003977773245424032, + 0.00447088573127985, + 0.006387628614902496, + 0.007339254952967167, + 0.004733953159302473, + 0.0034416085109114647, + 0.008870268240571022, + 0.0038452271837741137, + 0.002051738789305091, + 0.0037045273929834366, + 0.003305483376607299, + 0.005352950654923916, + 0.004210793878883123, + 0.0035707030910998583, + 0.005503732245415449, + 0.006104274187237024, + 0.006412046495825052, + 0.006772355642169714, + 0.006532082334160805, + 0.004976195283234119, + 0.0030791847966611385, + 0.0027544009499251842, + 0.0021928807254880667, + 0.0029901007656008005, + 0.004670525435358286, + 0.011231842450797558, + 0.004338476341217756, + 0.0060735843144357204, + 0.008009491488337517, + 0.008264646865427494, + 0.004905242472887039, + 0.008888292126357555, + 0.005716108251363039, + 0.007416355423629284, + 0.008065784350037575, + 0.008424204774200916, + 0.010790355503559113, + 0.008054210804402828, + 0.009078659117221832, + 0.0095960833132267, + 0.009445260278880596, + 0.00479111960157752, + 0.008619135245680809, + 0.010203415527939796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.04304387420415878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, uninterrupted tracking shot, maintaining an eye-level perspective with its subject, follows a paper airplane as it gracefully navigates a vast, desolate modern office. Meticulously folded from a single sheet of crisp white paper, its sharp creases forming the classic silhouette of a boomerang, the airplane glides silently and effortlessly through the still air. Cool, sterile light from long fluorescent lamps embedded in the ceiling bathes the scene, casting soft, elongated shadows and creating sharp, mirrored highlights on the polished concrete floors and smooth tabletops. The camera soars along a serene trajectory over rows of empty, minimalist desks, flying past the dark, reflective surfaces of powered-off computer screens and neatly arranged keyboards. Its gentle tilt, its uninterrupted motion contrasting with the static surroundings of abandoned ergonomic chairs, creates an atmosphere of profound tranquility and quiet solitude, a moment of simple, poetic beauty capturing the quiet heart of the corporate world.", + "seed": 14828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_7512_seed_14828_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": -0.08568112384527922, + "vlm_logprob": -0.08568112384527922, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.08568112384527922, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12037092447280884 + ], + "opposite_logprob_tokens": [ + -2.245370864868164 + ], + "logprob_reward": -0.12037092447280884, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12037092447280884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1384153813123703, + 0.26378655433654785, + 0.04819931089878082, + 0.07253167033195496, + 0.04603972285985947, + 0.1187746673822403, + 0.06839770078659058, + 0.055488571524620056, + 0.10098953545093536, + 0.0873769074678421 + ], + "credit_map": [ + 0.0162754338234663, + 0.015425155870616436, + 0.02789021097123623, + 0.016053413972258568, + 0.01968998648226261, + 0.019068434834480286, + 0.010975589975714684, + 0.01303714420646429, + 0.01649264059960842, + 0.01954696699976921, + 0.03868919610977173, + 0.03283023461699486, + 0.04638521745800972, + 0.022312868386507034, + 0.031788170337677, + 0.055741265416145325, + 0.0047887153923511505, + 0.0030706014949828386, + 0.00564238429069519, + 0.007824179716408253, + 0.004777759313583374, + 0.004830723628401756, + 0.0068047684617340565, + 0.01046018023043871, + 0.008815592154860497, + 0.010050017386674881, + 0.010022721253335476, + 0.012648491188883781, + 0.0051860567182302475, + 0.010806698352098465, + 0.007148394826799631, + 0.007853698916733265, + 0.004176300019025803, + 0.006578127387911081, + 0.009224562905728817, + 0.006750204134732485, + 0.004683203529566526, + 0.004444546531885862, + 0.005555677227675915, + 0.004627102520316839, + 0.012221517972648144, + 0.007434291765093803, + 0.017786996439099312, + 0.026901034638285637, + 0.008667844347655773, + 0.010998875834047794, + 0.021333986893296242, + 0.013430111110210419, + 0.01151074469089508, + 0.008853218518197536, + 0.00825512409210205, + 0.00729692867025733, + 0.008388045243918896, + 0.008842888288199902, + 0.007321154698729515, + 0.007929597049951553, + 0.0027819392271339893, + 0.0027863422874361277, + 0.01586146652698517, + 0.01800352707505226, + 0.0025437248405069113, + 0.0036683198995888233, + 0.0061765811406075954, + 0.003666665405035019, + 0.015556611120700836, + 0.008760916069149971, + 0.014341047033667564, + 0.014028842560946941, + 0.016261303797364235, + 0.010287859477102757, + 0.011177107691764832, + 0.010575844906270504, + 0.00914565660059452, + 0.009465161710977554, + 0.0077669937163591385, + 0.01313842087984085, + 0.007441041059792042, + 0.009256955236196518, + 0.022910702973604202, + 0.008251971565186977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.12037092447280884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990857362747192 + ], + "opposite_logprob_tokens": [ + -1.8049085140228271 + ], + "logprob_reward": -0.17990857362747192, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990857362747192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2197141945362091, + 0.4400154948234558, + 0.05440420284867287, + 0.04133003205060959, + 0.030758928507566452, + 0.046908821910619736, + 0.035144347697496414, + 0.03262806683778763, + 0.07199007272720337, + 0.027105851098895073 + ], + "credit_map": [ + 0.01872069202363491, + 0.028499966487288475, + 0.026269763708114624, + 0.027325846254825592, + 0.03407007455825806, + 0.035701848566532135, + 0.02128453180193901, + 0.027841469272971153, + 0.023526303470134735, + 0.025969533249735832, + 0.04705243557691574, + 0.07508630305528641, + 0.07227370887994766, + 0.02944156900048256, + 0.07352372258901596, + 0.09314187616109848, + 0.0049842107109725475, + 0.006819723639637232, + 0.005842937156558037, + 0.00624261936172843, + 0.008375353179872036, + 0.003647224046289921, + 0.007000720594078302, + 0.011491412296891212, + 0.004934700205922127, + 0.006714654620736837, + 0.004634056705981493, + 0.008956197649240494, + 0.002583899302408099, + 0.0037144790403544903, + 0.003909604158252478, + 0.005882443394511938, + 0.002078932011500001, + 0.0028061040211468935, + 0.009487784467637539, + 0.007403121795505285, + 0.0022579762153327465, + 0.0023675356060266495, + 0.0027626189403235912, + 0.0015948544023558497, + 0.0048071350902318954, + 0.003082206239923835, + 0.00650686863809824, + 0.007999172434210777, + 0.003782078390941024, + 0.004490713123232126, + 0.009706891141831875, + 0.006533753126859665, + 0.003414339618757367, + 0.002796910936012864, + 0.00530956732109189, + 0.005323995836079121, + 0.003636555280536413, + 0.003080823691561818, + 0.0064675589092075825, + 0.0051145958714187145, + 0.002545885043218732, + 0.0032806682866066694, + 0.004780560731887817, + 0.005839155986905098, + 0.0023455715272575617, + 0.006115184165537357, + 0.003288199892267585, + 0.004432843532413244, + 0.00534852547571063, + 0.006819804199039936, + 0.010449083521962166, + 0.009917988441884518, + 0.01209895871579647, + 0.006312989629805088, + 0.01234822440892458, + 0.008694500662386417, + 0.00298096239566803, + 0.002962794154882431, + 0.0028406805358827114, + 0.0050344280898571014, + 0.0023293106351047754, + 0.0024856121744960546, + 0.004828230012208223, + 0.0036438354291021824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.17990857362747192 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038089267909526825 + ], + "opposite_logprob_tokens": [ + -3.2880892753601074 + ], + "logprob_reward": -0.038089267909526825, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089267909526825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10577749460935593, + 0.2462334781885147, + 0.09252989292144775, + 0.09637504070997238, + 0.09579169005155563, + 0.07760759443044662, + 0.11197954416275024, + 0.06687694787979126, + 0.032777536660432816, + 0.07405070215463638 + ], + "credit_map": [ + 0.014025770127773285, + 0.014364823698997498, + 0.024309568107128143, + 0.01244824007153511, + 0.011773033067584038, + 0.013822264038026333, + 0.0071639372035861015, + 0.007869872264564037, + 0.019610246643424034, + 0.02433452010154724, + 0.04311316832900047, + 0.03820078447461128, + 0.02597898244857788, + 0.021729741245508194, + 0.030528049916028976, + 0.04273802414536476, + 0.013446638360619545, + 0.008924565277993679, + 0.009839560836553574, + 0.01916927471756935, + 0.008995728567242622, + 0.006780213676393032, + 0.013806200586259365, + 0.011567724868655205, + 0.0123535031452775, + 0.010563836432993412, + 0.014046652242541313, + 0.016971521079540253, + 0.0087959049269557, + 0.013488817028701305, + 0.010252370499074459, + 0.00990244373679161, + 0.009031732566654682, + 0.011410150676965714, + 0.017941433936357498, + 0.013808656483888626, + 0.010678611695766449, + 0.009565961547195911, + 0.014035073108971119, + 0.009320087730884552, + 0.006961653009057045, + 0.006128124427050352, + 0.011770334094762802, + 0.016850046813488007, + 0.0061670527793467045, + 0.008260699920356274, + 0.012042855843901634, + 0.009426837787032127, + 0.0181567445397377, + 0.013603129424154758, + 0.015283582732081413, + 0.011514606885612011, + 0.012544656172394753, + 0.008934438228607178, + 0.011262567713856697, + 0.020679833367466927, + 0.003907870966941118, + 0.005433589685708284, + 0.016543515026569366, + 0.02013379894196987, + 0.002577266190201044, + 0.006070224568247795, + 0.0057052369229495525, + 0.006505449768155813, + 0.0029395208694040775, + 0.003424700116738677, + 0.005140988621860743, + 0.004349042195826769, + 0.0046016862615942955, + 0.0028816768899559975, + 0.005132456310093403, + 0.004307472612708807, + 0.007543143350630999, + 0.010371224023401737, + 0.0062041329219937325, + 0.010942338034510612, + 0.005908553954213858, + 0.007560138124972582, + 0.017383845522999763, + 0.008137323893606663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.038089267909526825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084645055234432 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084645055234432, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084645055234432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13543447852134705, + 0.23595887422561646, + 0.2060030698776245, + 0.059485889971256256, + 0.02620561234652996, + 0.050342168658971786, + 0.03850318491458893, + 0.14882716536521912, + 0.0360264927148819, + 0.06321302801370621 + ], + "credit_map": [ + 0.015258384868502617, + 0.021195504814386368, + 0.03200606629252434, + 0.01511895377188921, + 0.01852734386920929, + 0.014195443131029606, + 0.009161751717329025, + 0.009971028193831444, + 0.01344409491866827, + 0.01314820908010006, + 0.03217078745365143, + 0.051339153200387955, + 0.02180938608944416, + 0.014444706030189991, + 0.042744845151901245, + 0.04685769975185394, + 0.022941119968891144, + 0.023615699261426926, + 0.023222506046295166, + 0.055901870131492615, + 0.013791762292385101, + 0.01501548197120428, + 0.025319717824459076, + 0.02619490772485733, + 0.005841339007019997, + 0.005044716876000166, + 0.009315608069300652, + 0.01510606613010168, + 0.00447703106328845, + 0.008157533593475819, + 0.006144147366285324, + 0.0053994497284293175, + 0.002681146375834942, + 0.0038049041759222746, + 0.005613546818494797, + 0.0038732988759875298, + 0.002175604458898306, + 0.0029549936298280954, + 0.0025866690557450056, + 0.002515449421480298, + 0.005441327579319477, + 0.00298407138325274, + 0.006892481818795204, + 0.010155715048313141, + 0.0035431943833827972, + 0.004811664577573538, + 0.009878359735012054, + 0.0066353497095406055, + 0.004645489621907473, + 0.0034564840607345104, + 0.005315394140779972, + 0.0052964044734835625, + 0.005857081618160009, + 0.004369445610791445, + 0.004585669841617346, + 0.004977213218808174, + 0.008568513207137585, + 0.006812376901507378, + 0.03602014482021332, + 0.05582467466592789, + 0.007006289437413216, + 0.012078305706381798, + 0.013975581154227257, + 0.008541279472410679, + 0.00377234211191535, + 0.0029650391079485416, + 0.006411712151020765, + 0.0048197475261986256, + 0.003861173987388611, + 0.00448474008589983, + 0.005378386937081814, + 0.004333354067057371, + 0.0077086035162210464, + 0.009394810535013676, + 0.005269367713481188, + 0.008747661486268044, + 0.005468151532113552, + 0.007072482258081436, + 0.013772540725767612, + 0.005779411643743515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.011084645055234432 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895220816135406 + ], + "opposite_logprob_tokens": [ + -2.5789523124694824 + ], + "logprob_reward": -0.07895220816135406, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895220816135406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10769620537757874, + 0.22133967280387878, + 0.08911696076393127, + 0.11683560162782669, + 0.07970704138278961, + 0.10848365724086761, + 0.07946180552244186, + 0.07825157791376114, + 0.04676477983593941, + 0.0723426416516304 + ], + "credit_map": [ + 0.012843355536460876, + 0.01406935602426529, + 0.019403748214244843, + 0.013718937523663044, + 0.016586650162935257, + 0.012524081394076347, + 0.008734703063964844, + 0.009815381839871407, + 0.01710495911538601, + 0.018848348408937454, + 0.017779620364308357, + 0.046387359499931335, + 0.030667655169963837, + 0.020628998056054115, + 0.03529836982488632, + 0.0346243642270565, + 0.00981361698359251, + 0.007326117251068354, + 0.010609133169054985, + 0.019966576248407364, + 0.009111102670431137, + 0.007241139188408852, + 0.01357182301580906, + 0.011477462947368622, + 0.0077591887675225735, + 0.011317918077111244, + 0.014488035812973976, + 0.01959673874080181, + 0.009713712148368359, + 0.02990601770579815, + 0.010996617376804352, + 0.01305738091468811, + 0.0054560150019824505, + 0.009677397087216377, + 0.015629928559064865, + 0.012762513943016529, + 0.008734655566513538, + 0.007411806844174862, + 0.012219035997986794, + 0.007815687917172909, + 0.01283645536750555, + 0.0073716710321605206, + 0.014191944152116776, + 0.020391954109072685, + 0.008955607190728188, + 0.012979020364582539, + 0.018580174073576927, + 0.01317683607339859, + 0.010662827640771866, + 0.007598394062370062, + 0.008924705907702446, + 0.009843623265624046, + 0.013920585624873638, + 0.007244439329952002, + 0.008603704161942005, + 0.012663532979786396, + 0.007829523645341396, + 0.00823521800339222, + 0.010855020023882389, + 0.013698205351829529, + 0.007278472650796175, + 0.008723522536456585, + 0.010327783413231373, + 0.011303838342428207, + 0.004929734393954277, + 0.0041750045493245125, + 0.009502853266894817, + 0.00599231431260705, + 0.005379696376621723, + 0.004669826477766037, + 0.006791701074689627, + 0.0053236521780490875, + 0.0076872739009559155, + 0.007108835503458977, + 0.006730739027261734, + 0.012323526665568352, + 0.006234045140445232, + 0.007694597821682692, + 0.017968250438570976, + 0.006595375947654247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.07895220816135406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, uninterrupted tracking shot, maintaining an eye-level perspective with its subject, follows a paper airplane as it gracefully navigates a vast, desolate modern office. Meticulously folded from a single sheet of crisp white paper, its sharp creases forming the classic silhouette of a boomerang, the airplane glides silently and effortlessly through the still air. Cool, sterile light from long fluorescent lamps embedded in the ceiling bathes the scene, casting soft, elongated shadows and creating sharp, mirrored highlights on the polished concrete floors and smooth tabletops. The camera soars along a serene trajectory over rows of empty, minimalist desks, flying past the dark, reflective surfaces of powered-off computer screens and neatly arranged keyboards. Its gentle tilt, its uninterrupted motion contrasting with the static surroundings of abandoned ergonomic chairs, creates an atmosphere of profound tranquility and quiet solitude, a moment of simple, poetic beauty capturing the quiet heart of the corporate world.", + "seed": 14829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_13/index_7512_seed_14829_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": -0.10328988768160344, + "vlm_logprob": -0.10328988768160344, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.10328988768160344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0741010457277298 + ], + "opposite_logprob_tokens": [ + -2.699100971221924 + ], + "logprob_reward": -0.0741010457277298, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0741010457277298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2868380844593048, + 0.08945796638727188, + 0.09722423553466797, + 0.09345994144678116, + 0.05433136597275734, + 0.05997756868600845, + 0.04635819047689438, + 0.022920828312635422, + 0.06744422018527985, + 0.18198762834072113 + ], + "credit_map": [ + 0.015052739530801773, + 0.022870052605867386, + 0.06047964096069336, + 0.02505519427359104, + 0.02175748720765114, + 0.014687597751617432, + 0.031089359894394875, + 0.0958460122346878, + 0.0068412418477237225, + 0.014497599564492702, + 0.008369329385459423, + 0.010070893913507462, + 0.013846305198967457, + 0.011436068452894688, + 0.014894668944180012, + 0.009501852095127106, + 0.011915084905922413, + 0.00852951593697071, + 0.010285270400345325, + 0.01694624312222004, + 0.008442090824246407, + 0.011632610112428665, + 0.017088240012526512, + 0.012385175563395023, + 0.011060459539294243, + 0.00967392884194851, + 0.012429160065948963, + 0.011953650042414665, + 0.010878254659473896, + 0.006902046035975218, + 0.016073500737547874, + 0.014488938264548779, + 0.008051451295614243, + 0.004049796611070633, + 0.00779716344550252, + 0.007624094840139151, + 0.006043562665581703, + 0.003740070154890418, + 0.00969716440886259, + 0.007328052073717117, + 0.00730088772252202, + 0.007254823576658964, + 0.007363986689597368, + 0.009354575537145138, + 0.00528584374114871, + 0.007799570448696613, + 0.007030282635241747, + 0.008587593212723732, + 0.004692529793828726, + 0.005894154775887728, + 0.007902391254901886, + 0.0036431928165256977, + 0.0032882075756788254, + 0.007046827115118504, + 0.007894506677985191, + 0.0059963795356452465, + 0.0019630989991128445, + 0.0018014822853729129, + 0.002676605014130473, + 0.003435899969190359, + 0.0031058548483997583, + 0.0026586782187223434, + 0.0029210879001766443, + 0.004358117934316397, + 0.007672965060919523, + 0.008765959180891514, + 0.009215567260980606, + 0.005941295996308327, + 0.00695083336904645, + 0.003949524834752083, + 0.01594126969575882, + 0.009006799198687077, + 0.009342680685222149, + 0.0036984989419579506, + 0.011273868381977081, + 0.008744804188609123, + 0.012328113429248333, + 0.007879920303821564, + 0.07084710150957108, + 0.05787263438105583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0741010457277298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134672939777374 + ], + "opposite_logprob_tokens": [ + -1.313467264175415 + ], + "logprob_reward": -0.3134672939777374, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134672939777374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34237247705459595, + 0.09148535877466202, + 0.10520859807729721, + 0.22269131243228912, + 0.03139963746070862, + 0.07900357246398926, + 0.01752473972737789, + 0.03765135258436203, + 0.03477376326918602, + 0.03788924962282181 + ], + "credit_map": [ + 0.039037659764289856, + 0.049379248172044754, + 0.029316779226064682, + 0.04660344496369362, + 0.03665030747652054, + 0.04661526903510094, + 0.046954330056905746, + 0.04781540483236313, + 0.006778862792998552, + 0.009031567722558975, + 0.012397781014442444, + 0.01174535695463419, + 0.008899467065930367, + 0.00963694415986538, + 0.015335132367908955, + 0.017660250887274742, + 0.010679489932954311, + 0.011849657632410526, + 0.008905193768441677, + 0.02100793831050396, + 0.008626466616988182, + 0.013155444525182247, + 0.01971772313117981, + 0.011266687884926796, + 0.028912438079714775, + 0.024759531021118164, + 0.03769580274820328, + 0.028053198009729385, + 0.02730921469628811, + 0.019009031355381012, + 0.031751807779073715, + 0.025200288742780685, + 0.0038105957210063934, + 0.002623156877234578, + 0.0042077526450157166, + 0.0042944010347127914, + 0.003794843563809991, + 0.002861032960936427, + 0.005128263495862484, + 0.004679590463638306, + 0.008345220237970352, + 0.005048188380897045, + 0.0076056793332099915, + 0.011692747473716736, + 0.004985249135643244, + 0.01051405444741249, + 0.016666769981384277, + 0.014145666733384132, + 0.0015921558951959014, + 0.0032996623776853085, + 0.0027181997429579496, + 0.0015486245974898338, + 0.0011984978336840868, + 0.002520914189517498, + 0.002509888494387269, + 0.0021367957815527916, + 0.004232574719935656, + 0.004153416492044926, + 0.00587192689999938, + 0.004155208822339773, + 0.005263681057840586, + 0.003993635065853596, + 0.0050695207901299, + 0.004911389667540789, + 0.0052681490778923035, + 0.003992559853941202, + 0.005829933565109968, + 0.00433554220944643, + 0.004808205645531416, + 0.002349681220948696, + 0.00462756073102355, + 0.003562128869816661, + 0.0035029887221753597, + 0.0027729233261197805, + 0.005259126424789429, + 0.004423667676746845, + 0.004516016226261854, + 0.0030720762442797422, + 0.008089661598205566, + 0.006252788472920656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.3134672939777374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865466430783272 + ], + "opposite_logprob_tokens": [ + -3.048654556274414 + ], + "logprob_reward": -0.04865466430783272, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865466430783272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26623380184173584, + 0.09082811325788498, + 0.040100354701280594, + 0.1006147712469101, + 0.09428530186414719, + 0.037132855504751205, + 0.03856971487402916, + 0.05075820907950401, + 0.059358395636081696, + 0.22211848199367523 + ], + "credit_map": [ + 0.020678222179412842, + 0.03832456097006798, + 0.02949630469083786, + 0.033856991678476334, + 0.029848547652363777, + 0.03272185102105141, + 0.03230135515332222, + 0.04900595545768738, + 0.007144775241613388, + 0.009634207934141159, + 0.009151392616331577, + 0.014406000263988972, + 0.00972207635641098, + 0.008591266348958015, + 0.022906219586730003, + 0.009272177703678608, + 0.004565114621073008, + 0.003974023275077343, + 0.003111387835815549, + 0.0060231345705688, + 0.0041593145579099655, + 0.00556006096303463, + 0.008242475800216198, + 0.004464844707399607, + 0.016601860523223877, + 0.009556259028613567, + 0.022817211225628853, + 0.010293720290064812, + 0.009560169652104378, + 0.008221326395869255, + 0.013077417388558388, + 0.010486804880201817, + 0.007574861869215965, + 0.005960467271506786, + 0.014416597783565521, + 0.013496438041329384, + 0.019082844257354736, + 0.00900180172175169, + 0.014748148620128632, + 0.01000414602458477, + 0.0056206476874649525, + 0.0032250939402729273, + 0.004456851165741682, + 0.003105465555563569, + 0.003120105480775237, + 0.005271256435662508, + 0.004777689464390278, + 0.00755574693903327, + 0.0033031979110091925, + 0.0052168541587889194, + 0.006439319346100092, + 0.0029555330984294415, + 0.0030998168513178825, + 0.005330886226147413, + 0.006711751688271761, + 0.005512355361133814, + 0.004381600301712751, + 0.003885620506480336, + 0.005451114848256111, + 0.0068012503907084465, + 0.007088693790137768, + 0.005763492546975613, + 0.006906108930706978, + 0.010480332188308239, + 0.007392804138362408, + 0.004678446799516678, + 0.008345036767423153, + 0.005009324289858341, + 0.007496302016079426, + 0.0045423987321555614, + 0.012591472826898098, + 0.009302614256739616, + 0.011399941518902779, + 0.006488016806542873, + 0.009721148759126663, + 0.05130843445658684, + 0.014026611112058163, + 0.007614391855895519, + 0.0869331806898117, + 0.03462676703929901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.04865466430783272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820111833512783 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.01820111833512783, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820111833512783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30713072419166565, + 0.10022047907114029, + 0.04593028873205185, + 0.2107555866241455, + 0.06010396033525467, + 0.030567720532417297, + 0.050772957503795624, + 0.05207902193069458, + 0.034917738288640976, + 0.10752154886722565 + ], + "credit_map": [ + 0.02068203128874302, + 0.022032544016838074, + 0.07492996007204056, + 0.033571720123291016, + 0.020354412496089935, + 0.02693159133195877, + 0.035392601042985916, + 0.07323585450649261, + 0.008541258051991463, + 0.012545271776616573, + 0.012181819416582584, + 0.019608529284596443, + 0.012082580476999283, + 0.008762851357460022, + 0.019361071288585663, + 0.0071371011435985565, + 0.004804762080311775, + 0.004400166682898998, + 0.0032555018551647663, + 0.007033813279122114, + 0.0047283051535487175, + 0.007228227332234383, + 0.009108901023864746, + 0.005370608996599913, + 0.03173540160059929, + 0.020664317533373833, + 0.03820880502462387, + 0.02056351862847805, + 0.028312524780631065, + 0.018863286823034286, + 0.028645813465118408, + 0.023761922493577003, + 0.007077435962855816, + 0.005778768099844456, + 0.009007284417748451, + 0.008317440748214722, + 0.008317410945892334, + 0.006874663755297661, + 0.008304859511554241, + 0.006426092702895403, + 0.002506533870473504, + 0.00207349075935781, + 0.0038819597102701664, + 0.005184956826269627, + 0.0017113516805693507, + 0.00453138304874301, + 0.005780710838735104, + 0.004897334147244692, + 0.0050500123761594296, + 0.008512287400662899, + 0.008260156027972698, + 0.0041871522553265095, + 0.0034679442178457975, + 0.007236336823552847, + 0.007274571806192398, + 0.006784494966268539, + 0.006275145336985588, + 0.0049154977314174175, + 0.007828234694898129, + 0.005365696270018816, + 0.007810153998434544, + 0.0055254725739359856, + 0.008003730326890945, + 0.0063550942577421665, + 0.0028014956042170525, + 0.004214722663164139, + 0.005321758333593607, + 0.0056822095066308975, + 0.004944919608533382, + 0.0032822478096932173, + 0.0037688561715185642, + 0.00490152882412076, + 0.005489124450832605, + 0.002786745550110936, + 0.006713132839649916, + 0.004834438674151897, + 0.007522395811975002, + 0.006833043415099382, + 0.03821784630417824, + 0.03512481227517128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.01820111833512783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a single sheet of crisp white paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane shaped like a boomerang with sharp creases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast, desolate modern office with empty minimalist desks and powered-off computer screens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting cool and sterile, coming from long fluorescent lamps and casting elongated shadows on polished concrete floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a fluid, uninterrupted tracking shot that soars at eye-level along a serene trajectory?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062025316059589386 + ], + "opposite_logprob_tokens": [ + -2.812025308609009 + ], + "logprob_reward": -0.062025316059589386, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062025316059589386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1769483983516693, + 0.06717420369386673, + 0.04305720329284668, + 0.30144062638282776, + 0.046452950686216354, + 0.05116511136293411, + 0.12559764087200165, + 0.06747424602508545, + 0.021600328385829926, + 0.09908927232027054 + ], + "credit_map": [ + 0.016555719077587128, + 0.019000845029950142, + 0.02691260725259781, + 0.028128977864980698, + 0.01416954305022955, + 0.018117493018507957, + 0.02099163644015789, + 0.03307158127427101, + 0.008332044817507267, + 0.008743131533265114, + 0.008815242908895016, + 0.007381509989500046, + 0.009626413695514202, + 0.009922731667757034, + 0.008280147798359394, + 0.006072978489100933, + 0.005534611642360687, + 0.004137204494327307, + 0.004012219607830048, + 0.005418929271399975, + 0.004624989815056324, + 0.006504741031676531, + 0.007490318268537521, + 0.005334187764674425, + 0.052958209067583084, + 0.028008634224534035, + 0.07201128453016281, + 0.030093785375356674, + 0.020897720009088516, + 0.025256257504224777, + 0.03775402158498764, + 0.03446070849895477, + 0.0053548929281532764, + 0.004334641620516777, + 0.006306922994554043, + 0.005650087725371122, + 0.005699742119759321, + 0.0043835206888616085, + 0.008089762181043625, + 0.006633380893617868, + 0.005871864501386881, + 0.005418007727712393, + 0.005775778088718653, + 0.006817006506025791, + 0.004551928956061602, + 0.006886068731546402, + 0.006106511689722538, + 0.009737947024405003, + 0.011777631938457489, + 0.025213899090886116, + 0.018955733627080917, + 0.009015630930662155, + 0.007807101588696241, + 0.01625635288655758, + 0.022885728627443314, + 0.013685578480362892, + 0.007044096942991018, + 0.005805106833577156, + 0.006606690585613251, + 0.009498007595539093, + 0.008855706080794334, + 0.007664755918085575, + 0.007601874880492687, + 0.014398010447621346, + 0.002867711940780282, + 0.002662941114977002, + 0.0029892052989453077, + 0.0018274568719789386, + 0.002997130621224642, + 0.002530860248953104, + 0.002989063737913966, + 0.0027359577361494303, + 0.006111036520451307, + 0.0031540184281766415, + 0.005366147495806217, + 0.012003923766314983, + 0.006410655099898577, + 0.005657997447997332, + 0.035584401339292526, + 0.024801086634397507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.062025316059589386 + } + ], + "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/0000110/group_14/index_tensor([3660, 3660])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_14/index_tensor([3660, 3660])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..82a4217a41633cf1189a38560cbfb74ecbc7c503 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_14/index_tensor([3660, 3660])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, cinematic close-up, with a slow, almost imperceptible pan to the right, closely follows an Australian Gila monster as it slowly searches for water in the dry, cracked red earth. The sharpened image exaggerates the texture of the lizard's rough, sandpaper-like skin, covered in sharp, keratinized, conical spines. The edges and textures of each scale are clearly visible, and its intricate camouflage pattern of desert ochre, sandy yellow, and dark brown appears strikingly realistic under the blazing afternoon sun. It advances with a measured, stop-and-go rhythm, cautiously lifting a foot, pausing briefly in mid-air before firmly lowering it. Its small, obsidian eyes flick vigilantly, and its head moves slightly from side to side, as if searching for water. Subtle film grain dances across the frame, combined with a noticeable high-noise texture, imbuing this leisurely, focused scene with a powerful documentary quality, resembling a moment brimming with raw vitality captured from a top-tier nature documentary.", + "seed": 982, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_14/index_3660_seed_982_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": -0.11411135606467723, + "vlm_logprob": -0.11411135606467723, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.11411135606467723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.23164035379886627 + ], + "opposite_logprob_tokens": [ + -1.606640338897705 + ], + "logprob_reward": -0.23164035379886627, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23164035379886627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2507924437522888, + 0.05281737074255943, + 0.04287335276603699, + 0.05986795946955681, + 0.2371334433555603, + 0.07599333673715591, + 0.08950138092041016, + 0.06087121367454529, + 0.05908558890223503, + 0.07106383889913559 + ], + "credit_map": [ + 0.045364055782556534, + 0.013261807151138783, + 0.03178904578089714, + 0.01769973710179329, + 0.01917251944541931, + 0.023539481684565544, + 0.06423011422157288, + 0.03573570400476456, + 0.009023742750287056, + 0.0034371709916740656, + 0.014890824444591999, + 0.005445034708827734, + 0.003626821795478463, + 0.004936308600008488, + 0.0049826353788375854, + 0.006474839523434639, + 0.007678190246224403, + 0.004641546402126551, + 0.005428445525467396, + 0.007600617595016956, + 0.004937268793582916, + 0.004574658814817667, + 0.0050973594188690186, + 0.002915268996730447, + 0.007181093562394381, + 0.006925488822162151, + 0.005989885423332453, + 0.006666246335953474, + 0.009310686029493809, + 0.007815483026206493, + 0.00830500666052103, + 0.007674072869122028, + 0.0659911185503006, + 0.03287371248006821, + 0.02420695312321186, + 0.018571140244603157, + 0.01619827374815941, + 0.03212510421872139, + 0.01769917830824852, + 0.029467975720763206, + 0.009705893695354462, + 0.009019206278026104, + 0.008011546917259693, + 0.011305094696581364, + 0.00733130844309926, + 0.013754008337855339, + 0.009301906451582909, + 0.0075643714517354965, + 0.008324852213263512, + 0.012969285249710083, + 0.00855297315865755, + 0.011612124741077423, + 0.008985438384115696, + 0.012135423719882965, + 0.01102113351225853, + 0.01590016484260559, + 0.005228021647781134, + 0.004887237213551998, + 0.007144193165004253, + 0.006293799728155136, + 0.010122695937752724, + 0.012046253308653831, + 0.004845260176807642, + 0.010303757153451443, + 0.006154954899102449, + 0.009256294928491116, + 0.007317381910979748, + 0.005847879219800234, + 0.014234762638807297, + 0.004863313399255276, + 0.006345952861011028, + 0.0050650532357394695, + 0.00839394424110651, + 0.008760537020862103, + 0.007793658412992954, + 0.010723183862864971, + 0.007535367272794247, + 0.01152786798775196, + 0.009017071686685085, + 0.007312210276722908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.23164035379886627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.018216218799352646 + ], + "opposite_logprob_tokens": [ + -4.018216133117676 + ], + "logprob_reward": -0.018216218799352646, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018216218799352646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30556145310401917, + 0.034561026841402054, + 0.058120112866163254, + 0.031013494357466698, + 0.11945022642612457, + 0.18380221724510193, + 0.07654442638158798, + 0.031646210700273514, + 0.039074551314115524, + 0.12022630870342255 + ], + "credit_map": [ + 0.07219020277261734, + 0.027603283524513245, + 0.03887521103024483, + 0.016905544325709343, + 0.027198724448680878, + 0.018895534798502922, + 0.07184294611215591, + 0.03204998001456261, + 0.004145515616983175, + 0.00374534516595304, + 0.006322485860437155, + 0.00559109915047884, + 0.003337165340781212, + 0.002849967684596777, + 0.003941555507481098, + 0.0046278913505375385, + 0.010668961331248283, + 0.005727526266127825, + 0.007269341964274645, + 0.010713201947510242, + 0.006655855569988489, + 0.007342138793319464, + 0.005490294191986322, + 0.0042527890764176846, + 0.0048618800938129425, + 0.0032740591559559107, + 0.0027701742947101593, + 0.0034180188085883856, + 0.003446550341323018, + 0.0049209739081561565, + 0.0038451519794762135, + 0.004476682748645544, + 0.023655056953430176, + 0.008723735809326172, + 0.007800583261996508, + 0.04589974135160446, + 0.005573073867708445, + 0.007583399768918753, + 0.01299985870718956, + 0.007214767392724752, + 0.018551236018538475, + 0.013397310860455036, + 0.0560564361512661, + 0.015015549026429653, + 0.010922892019152641, + 0.0058611370623111725, + 0.05401110649108887, + 0.009986555203795433, + 0.014390883035957813, + 0.007286483887583017, + 0.00930572859942913, + 0.01048379298299551, + 0.008272087201476097, + 0.011939989402890205, + 0.007373597472906113, + 0.007491853553801775, + 0.0021257828921079636, + 0.0036357701756060123, + 0.0038746288046240807, + 0.003930404782295227, + 0.003837552620097995, + 0.0078092291951179504, + 0.0036954304669052362, + 0.0027374099008738995, + 0.003317373339086771, + 0.004584033973515034, + 0.005207370035350323, + 0.003558612195774913, + 0.007414375897496939, + 0.0035294711124151945, + 0.006445717066526413, + 0.005017593037337065, + 0.012147956527769566, + 0.016140442341566086, + 0.01306142471730709, + 0.017306489869952202, + 0.014135553501546383, + 0.021258728578686714, + 0.014463829807937145, + 0.011711874976754189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.018216218799352646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.016083383932709694 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.016083383932709694, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083383932709694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12281765043735504, + 0.07197631150484085, + 0.05153052508831024, + 0.05422814562916756, + 0.23397834599018097, + 0.21052023768424988, + 0.06856320053339005, + 0.06179128214716911, + 0.05979250371456146, + 0.06480183452367783 + ], + "credit_map": [ + 0.028109507635235786, + 0.011050348170101643, + 0.015358518809080124, + 0.0059180669486522675, + 0.010372989811003208, + 0.007710738107562065, + 0.03056303784251213, + 0.013734436593949795, + 0.012597375549376011, + 0.006594279780983925, + 0.015604466199874878, + 0.008972283452749252, + 0.0069283549673855305, + 0.0067901331931352615, + 0.008139451965689659, + 0.00634996360167861, + 0.008317624218761921, + 0.005949106067419052, + 0.009976035915315151, + 0.007553720846772194, + 0.006086007691919804, + 0.004301105160266161, + 0.00525070633739233, + 0.0040962183848023415, + 0.004644433036446571, + 0.009958334267139435, + 0.004071021918207407, + 0.004739693831652403, + 0.01341341808438301, + 0.004585423972457647, + 0.006521993782371283, + 0.006293823942542076, + 0.02748301438987255, + 0.011600439436733723, + 0.012010786682367325, + 0.08794999867677689, + 0.009789601899683475, + 0.05042487382888794, + 0.02372928149998188, + 0.010990344919264317, + 0.021565239876508713, + 0.016955344006419182, + 0.06118527427315712, + 0.018415581434965134, + 0.011741734109818935, + 0.00927078165113926, + 0.05981341749429703, + 0.011572862043976784, + 0.008175262250006199, + 0.00801108404994011, + 0.006381693296134472, + 0.009229104965925217, + 0.00798364169895649, + 0.008873446844518185, + 0.008760771714150906, + 0.011148194782435894, + 0.005268924403935671, + 0.010033776052296162, + 0.005640539340674877, + 0.007117395289242268, + 0.00912002008408308, + 0.009540659375488758, + 0.007397163659334183, + 0.0076728020794689655, + 0.006223408039659262, + 0.008369985967874527, + 0.004961285274475813, + 0.0062823849730193615, + 0.013353616930544376, + 0.005076687317341566, + 0.009614253416657448, + 0.005910881794989109, + 0.007124373689293861, + 0.006768000312149525, + 0.007346887607127428, + 0.012394534423947334, + 0.006534439977258444, + 0.011089319363236427, + 0.007496494334191084, + 0.006047787610441446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.016083383932709694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.023304030299186707 + ], + "opposite_logprob_tokens": [ + -3.773303985595703 + ], + "logprob_reward": -0.023304030299186707, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304030299186707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05833740159869194, + 0.028290312737226486, + 0.020120181143283844, + 0.04526194930076599, + 0.10912935435771942, + 0.5364358425140381, + 0.07715775072574615, + 0.04275146871805191, + 0.02965834178030491, + 0.05285739526152611 + ], + "credit_map": [ + 0.009618179872632027, + 0.005408724304288626, + 0.009080401621758938, + 0.005799359176307917, + 0.007517754100263119, + 0.006221296731382608, + 0.008629782125353813, + 0.006061903666704893, + 0.005090105812996626, + 0.003196269040927291, + 0.0073106600902974606, + 0.0025466633960604668, + 0.003770096693187952, + 0.002226676791906357, + 0.001953867031261325, + 0.0021959745790809393, + 0.002417364390566945, + 0.002736482070758939, + 0.001906324177980423, + 0.003723845584318042, + 0.0029210669454187155, + 0.0016956243198364973, + 0.0031418385915458202, + 0.0015776342479512095, + 0.004414653405547142, + 0.006682246457785368, + 0.003963053226470947, + 0.005669321399182081, + 0.008493415080010891, + 0.004257711581885815, + 0.004751462489366531, + 0.007030084263533354, + 0.01501498930156231, + 0.008448194712400436, + 0.00768049294129014, + 0.028646251186728477, + 0.007871335372328758, + 0.010542919859290123, + 0.009067146107554436, + 0.02185802347958088, + 0.056825313717126846, + 0.04336823150515556, + 0.15477341413497925, + 0.04606863483786583, + 0.029422994703054428, + 0.020287280902266502, + 0.15652009844779968, + 0.029169881716370583, + 0.010194906033575535, + 0.008775380440056324, + 0.010658926330506802, + 0.01120066735893488, + 0.0067404769361019135, + 0.007728402502834797, + 0.012613245286047459, + 0.00924574676901102, + 0.004848272539675236, + 0.005387918557971716, + 0.005314795766025782, + 0.003923818003386259, + 0.004602835979312658, + 0.00826531182974577, + 0.005134205333888531, + 0.00527431583032012, + 0.0030800148379057646, + 0.004456397145986557, + 0.003903367556631565, + 0.005531725473701954, + 0.004250467289239168, + 0.002383620711043477, + 0.0037880714517086744, + 0.0022646801080554724, + 0.007023717276751995, + 0.0076020159758627415, + 0.00489015830680728, + 0.004600028973072767, + 0.010151217691600323, + 0.006777231115847826, + 0.006426336243748665, + 0.005386686883866787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.023304030299186707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.2813127934932709 + ], + "opposite_logprob_tokens": [ + -1.4063128232955933 + ], + "logprob_reward": -0.2813127934932709, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813127934932709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07438776642084122, + 0.06282559782266617, + 0.0348360538482666, + 0.02539619244635105, + 0.5168891549110413, + 0.09765986353158951, + 0.07082705944776535, + 0.05300421640276909, + 0.03144175559282303, + 0.032732412219047546 + ], + "credit_map": [ + 0.016794444993138313, + 0.005379780195653439, + 0.02048403024673462, + 0.005469175986945629, + 0.008547305129468441, + 0.006229541730135679, + 0.006803225260227919, + 0.00468025915324688, + 0.009668005630373955, + 0.0056590549647808075, + 0.014910480007529259, + 0.00877317413687706, + 0.004524593241512775, + 0.005066827405244112, + 0.007122261915355921, + 0.007101194933056831, + 0.00467279925942421, + 0.0033913860097527504, + 0.004441278520971537, + 0.006657733581960201, + 0.004059432540088892, + 0.0038912934251129627, + 0.004107775632292032, + 0.003614352783188224, + 0.0027196635492146015, + 0.004082558210939169, + 0.0022429509554058313, + 0.002755874302238226, + 0.004873144440352917, + 0.002557489089667797, + 0.0030083737801760435, + 0.0031561388168483973, + 0.0811077356338501, + 0.03062112256884575, + 0.02970152348279953, + 0.22066718339920044, + 0.01958518475294113, + 0.025814615190029144, + 0.03743699565529823, + 0.07195476442575455, + 0.009413846768438816, + 0.008921772241592407, + 0.02892405167222023, + 0.008127848617732525, + 0.006746833212673664, + 0.005349589977413416, + 0.023542853072285652, + 0.00663306238129735, + 0.008034542202949524, + 0.012555786408483982, + 0.007285909727215767, + 0.009497195482254028, + 0.005879039876163006, + 0.011048768647015095, + 0.008622986264526844, + 0.007902825251221657, + 0.004428348038345575, + 0.009420150890946388, + 0.0038125619757920504, + 0.00943052303045988, + 0.008114784955978394, + 0.006710059940814972, + 0.0068474323488771915, + 0.004240353126078844, + 0.0035213218070566654, + 0.004176002461463213, + 0.003743801498785615, + 0.004568030126392841, + 0.006564220879226923, + 0.00248789694160223, + 0.003241624915972352, + 0.003138852771371603, + 0.005345269571989775, + 0.0036725408863276243, + 0.003161402652040124, + 0.003705230075865984, + 0.0038614633958786726, + 0.004038822837173939, + 0.004234426189213991, + 0.0047132596373558044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.2813127934932709 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, cinematic close-up, with a slow, almost imperceptible pan to the right, closely follows an Australian Gila monster as it slowly searches for water in the dry, cracked red earth. The sharpened image exaggerates the texture of the lizard's rough, sandpaper-like skin, covered in sharp, keratinized, conical spines. The edges and textures of each scale are clearly visible, and its intricate camouflage pattern of desert ochre, sandy yellow, and dark brown appears strikingly realistic under the blazing afternoon sun. It advances with a measured, stop-and-go rhythm, cautiously lifting a foot, pausing briefly in mid-air before firmly lowering it. Its small, obsidian eyes flick vigilantly, and its head moves slightly from side to side, as if searching for water. Subtle film grain dances across the frame, combined with a noticeable high-noise texture, imbuing this leisurely, focused scene with a powerful documentary quality, resembling a moment brimming with raw vitality captured from a top-tier nature documentary.", + "seed": 983, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_14/index_3660_seed_983_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": -0.10461162263527513, + "vlm_logprob": -0.10461162263527513, + "vlm_yes_no_margin": 3.024999976158142, + "vlm_token_credit": -0.10461162263527513, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.14665587246418 + ], + "opposite_logprob_tokens": [ + -2.021655797958374 + ], + "logprob_reward": -0.14665587246418, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14665587246418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06040244176983833, + 0.22448104619979858, + 0.054685357958078384, + 0.11620193719863892, + 0.3404829800128937, + 0.07590464502573013, + 0.04093261808156967, + 0.02010980248451233, + 0.03400687128305435, + 0.03279227018356323 + ], + "credit_map": [ + 0.012584415264427662, + 0.007654625456780195, + 0.00922291912138462, + 0.00772927887737751, + 0.007498905528336763, + 0.00593290152028203, + 0.005486099049448967, + 0.004293297883123159, + 0.012662211433053017, + 0.004463682882487774, + 0.08124492317438126, + 0.00597215024754405, + 0.014683656394481659, + 0.09495062381029129, + 0.005670513492077589, + 0.00483326893299818, + 0.0052194418385624886, + 0.008766566403210163, + 0.008145375177264214, + 0.007326320745050907, + 0.005844174884259701, + 0.010602041147649288, + 0.0050123874098062515, + 0.0037690496537834406, + 0.03408019617199898, + 0.011955750174820423, + 0.01292336918413639, + 0.00942292157560587, + 0.011362341232597828, + 0.008736557327210903, + 0.00864920299500227, + 0.019071606919169426, + 0.05434595048427582, + 0.020227115601301193, + 0.13076142966747284, + 0.03683256730437279, + 0.014506570063531399, + 0.029504647478461266, + 0.017972877249121666, + 0.036331843584775925, + 0.007840943522751331, + 0.010598471388220787, + 0.0097215436398983, + 0.007146062795072794, + 0.011948312632739544, + 0.012013925239443779, + 0.00745348259806633, + 0.00918190460652113, + 0.004408327862620354, + 0.010671896860003471, + 0.004203799180686474, + 0.005720512475818396, + 0.006509603466838598, + 0.004334569443017244, + 0.0028006574138998985, + 0.0022832511458545923, + 0.0024547805078327656, + 0.0029661129228770733, + 0.0028530231211334467, + 0.002490746323019266, + 0.0027035672683268785, + 0.0023361877538263798, + 0.0018738927319645882, + 0.0024314909242093563, + 0.006854079198092222, + 0.004374986048787832, + 0.004221739713102579, + 0.0032599077094346285, + 0.003987213131040335, + 0.004168319050222635, + 0.00418993690982461, + 0.0029506897553801537, + 0.0055572157725691795, + 0.003952596802264452, + 0.0033119572326540947, + 0.0027341581881046295, + 0.005106596741825342, + 0.00452100345864892, + 0.004377748351544142, + 0.0032309929374605417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.14665587246418 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.005968604702502489 + ], + "opposite_logprob_tokens": [ + -5.1309685707092285 + ], + "logprob_reward": -0.005968604702502489, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005968604702502489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10804010927677155, + 0.5786283016204834, + 0.03129102289676666, + 0.04315290227532387, + 0.034584999084472656, + 0.03780985623598099, + 0.03085974231362343, + 0.06137997657060623, + 0.020988401025533676, + 0.05326478183269501 + ], + "credit_map": [ + 0.023099809885025024, + 0.016010502353310585, + 0.0186525397002697, + 0.012635053135454655, + 0.01193211879581213, + 0.009850204922258854, + 0.008726166561245918, + 0.007133709732443094, + 0.030515378341078758, + 0.009242202155292034, + 0.2148415744304657, + 0.016346173360943794, + 0.03613939508795738, + 0.24671649932861328, + 0.013322616927325726, + 0.011504454538226128, + 0.0036824941635131836, + 0.002297342522069812, + 0.011078918352723122, + 0.001986066810786724, + 0.0026003436651080847, + 0.004603879991918802, + 0.002833425533026457, + 0.002208553021773696, + 0.012173241004347801, + 0.00458189845085144, + 0.00452865194529295, + 0.0037134597077965736, + 0.004283232148736715, + 0.004175644367933273, + 0.0031156425829976797, + 0.006581132300198078, + 0.005639855284243822, + 0.002292670076712966, + 0.01224589254707098, + 0.003621528623625636, + 0.0016672074561938643, + 0.0033140182495117188, + 0.0020089121535420418, + 0.003794911317527294, + 0.006259996443986893, + 0.0044343373738229275, + 0.0047408235259354115, + 0.004376741126179695, + 0.004834380000829697, + 0.003952675964683294, + 0.00438713189214468, + 0.004823766183108091, + 0.003805988933891058, + 0.007811782415956259, + 0.0035466940607875586, + 0.0025070516858249903, + 0.004787496756762266, + 0.003945684991776943, + 0.0021627910900861025, + 0.0022922533098608255, + 0.0073251607827842236, + 0.00676997285336256, + 0.01324017345905304, + 0.006582242902368307, + 0.004059490282088518, + 0.0076536922715604305, + 0.005310634151101112, + 0.010438607074320316, + 0.0030630878172814846, + 0.0019202649127691984, + 0.00448156613856554, + 0.0013834077399224043, + 0.0019097639014944434, + 0.0029227223712950945, + 0.0023052655160427094, + 0.003002322744578123, + 0.005565050058066845, + 0.004973100032657385, + 0.006158827338367701, + 0.006433840375393629, + 0.006749479565769434, + 0.006485213525593281, + 0.007846962660551071, + 0.009052310138940811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.005968604702502489 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.023297391831874847 + ], + "opposite_logprob_tokens": [ + -3.7732973098754883 + ], + "logprob_reward": -0.023297391831874847, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297391831874847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04841695353388786, + 0.5206547975540161, + 0.10342376679182053, + 0.06265415996313095, + 0.07942654937505722, + 0.036356907337903976, + 0.038144588470458984, + 0.022463403642177582, + 0.0426739864051342, + 0.045784901827573776 + ], + "credit_map": [ + 0.007036943454295397, + 0.0068138069473207, + 0.0076792798936367035, + 0.006572361569851637, + 0.0073432014323771, + 0.005102480761706829, + 0.004069801419973373, + 0.0037990775890648365, + 0.027783913537859917, + 0.010343918576836586, + 0.1891627162694931, + 0.015310625545680523, + 0.0345054529607296, + 0.21974124014377594, + 0.012162008322775364, + 0.011644937098026276, + 0.012185134924948215, + 0.015774991363286972, + 0.013974495232105255, + 0.017788182944059372, + 0.011170120909810066, + 0.012236757203936577, + 0.013971660286188126, + 0.006322420667856932, + 0.013057257048785686, + 0.005322772078216076, + 0.005227557849138975, + 0.006018773652613163, + 0.00793883390724659, + 0.011587847024202347, + 0.00628710025921464, + 0.007214020937681198, + 0.012511014007031918, + 0.004629114642739296, + 0.029901662841439247, + 0.008496761322021484, + 0.003571947105228901, + 0.007193221244961023, + 0.004151692148298025, + 0.008971137925982475, + 0.003812092123553157, + 0.00483045494183898, + 0.005202615633606911, + 0.0036652886774390936, + 0.004953425843268633, + 0.006501493044197559, + 0.0030372412875294685, + 0.004354297649115324, + 0.0060353390872478485, + 0.004507026635110378, + 0.005466544535011053, + 0.005020520184189081, + 0.004992393776774406, + 0.0054609207436442375, + 0.003302986267954111, + 0.003358859336003661, + 0.003119338070973754, + 0.004010287579149008, + 0.0024715519975870848, + 0.0024539027363061905, + 0.0026918661314994097, + 0.002337890211492777, + 0.0022559126373380423, + 0.0031226540450006723, + 0.008446969091892242, + 0.004710362292826176, + 0.00445941649377346, + 0.005140862427651882, + 0.005061545874923468, + 0.00543545326218009, + 0.005019218660891056, + 0.004400154110044241, + 0.012265457771718502, + 0.0062110451981425285, + 0.00414141733199358, + 0.0036428433377295732, + 0.005349442828446627, + 0.0068783024325966835, + 0.003681593108922243, + 0.0036148028448224068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.023297391831874847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.033713892102241516 + ], + "opposite_logprob_tokens": [ + -3.4087138175964355 + ], + "logprob_reward": -0.033713892102241516, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033713892102241516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05303262546658516, + 0.5036448240280151, + 0.1306808441877365, + 0.04195943474769592, + 0.0810757502913475, + 0.0745772272348404, + 0.050602272152900696, + 0.016377614811062813, + 0.02321699447929859, + 0.02483244612812996 + ], + "credit_map": [ + 0.009683912619948387, + 0.00927384290844202, + 0.007673860993236303, + 0.005337650887668133, + 0.00584535812959075, + 0.005467914044857025, + 0.005746311042457819, + 0.004003774840384722, + 0.02696729078888893, + 0.00728589715436101, + 0.18796029686927795, + 0.013638573698699474, + 0.03182952478528023, + 0.21259063482284546, + 0.01274716854095459, + 0.010625422932207584, + 0.014245490543544292, + 0.008200077340006828, + 0.048938047140836716, + 0.008044878952205181, + 0.010569303296506405, + 0.02452068217098713, + 0.010762417688965797, + 0.0053999400697648525, + 0.010602962225675583, + 0.004264850169420242, + 0.004857270047068596, + 0.00358755886554718, + 0.003590822219848633, + 0.006056873127818108, + 0.003557620570063591, + 0.005441479850560427, + 0.013259069994091988, + 0.00533175328746438, + 0.02932611294090748, + 0.010291132144629955, + 0.005415491294115782, + 0.006363540887832642, + 0.0044870213605463505, + 0.0066016255877912045, + 0.0062944781966507435, + 0.010473130270838737, + 0.010062338784337044, + 0.008136671036481857, + 0.011722132563591003, + 0.011608508415520191, + 0.0076422737911343575, + 0.008637689054012299, + 0.006489933002740145, + 0.006464111153036356, + 0.00808785855770111, + 0.009441923350095749, + 0.00462545920163393, + 0.006505974102765322, + 0.003935324493795633, + 0.005051686428487301, + 0.0025796822737902403, + 0.0019446860533207655, + 0.0027774940244853497, + 0.0020536051597446203, + 0.001778993639163673, + 0.001665002084337175, + 0.0018075768603011966, + 0.0017705762293189764, + 0.003548654494807124, + 0.0034648869186639786, + 0.0027469205670058727, + 0.0023391207214444876, + 0.0022755051031708717, + 0.002751791849732399, + 0.002795711625367403, + 0.0032944022677838802, + 0.0040761190466582775, + 0.002332626609131694, + 0.0028721794951707125, + 0.0025013452395796776, + 0.0033321459777653217, + 0.0034603187814354897, + 0.003031935077160597, + 0.0032257791608572006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.033713892102241516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.3134223520755768 + ], + "opposite_logprob_tokens": [ + -1.3134223222732544 + ], + "logprob_reward": -0.3134223520755768, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134223520755768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035042937844991684, + 0.3878940939903259, + 0.090349942445755, + 0.27669233083724976, + 0.04344331473112106, + 0.06185200810432434, + 0.02833598107099533, + 0.009265250526368618, + 0.021678922697901726, + 0.04544531926512718 + ], + "credit_map": [ + 0.0061926571652293205, + 0.0037766729947179556, + 0.0055167656391859055, + 0.0030263401567935944, + 0.0047169337049126625, + 0.00530365901067853, + 0.003971170634031296, + 0.002538736443966627, + 0.024541599676012993, + 0.009124304167926311, + 0.13898515701293945, + 0.01150443684309721, + 0.02359195426106453, + 0.16198191046714783, + 0.009910871274769306, + 0.008253834210336208, + 0.0077773951925337315, + 0.006529540289193392, + 0.03107992187142372, + 0.005633245222270489, + 0.00824520643800497, + 0.01966105028986931, + 0.0077528925612568855, + 0.003670676611363888, + 0.08548702299594879, + 0.029535729438066483, + 0.029632333666086197, + 0.02113243378698826, + 0.027690842747688293, + 0.02310405857861042, + 0.017778152599930763, + 0.042331743985414505, + 0.006661554798483849, + 0.0031100537162274122, + 0.01537399459630251, + 0.005719673819839954, + 0.0030101798474788666, + 0.003539248602464795, + 0.002519129076972604, + 0.0035094742197543383, + 0.007786193862557411, + 0.005741229746490717, + 0.010702532716095448, + 0.010329375974833965, + 0.007386384066194296, + 0.005976385436952114, + 0.006339304614812136, + 0.0075905886478722095, + 0.003700754838064313, + 0.004361601080745459, + 0.003055688925087452, + 0.004455877002328634, + 0.0038247578777372837, + 0.003162981476634741, + 0.002964767161756754, + 0.0028095499146729708, + 0.001135287107899785, + 0.0010532616870477796, + 0.0013194234343245625, + 0.0013377733994275331, + 0.0009891586378216743, + 0.0010090377181768417, + 0.0010100482031702995, + 0.001411259057931602, + 0.005426392890512943, + 0.0028637582436203957, + 0.0025789563078433275, + 0.0021021589636802673, + 0.0018403228605166078, + 0.003116052132099867, + 0.0017022235551849008, + 0.0020490551833063364, + 0.009657418355345726, + 0.005809823051095009, + 0.004335930570960045, + 0.0042793345637619495, + 0.006616925820708275, + 0.006352726370096207, + 0.004817765671759844, + 0.003575387643650174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.3134223520755768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, cinematic close-up, with a slow, almost imperceptible pan to the right, closely follows an Australian Gila monster as it slowly searches for water in the dry, cracked red earth. The sharpened image exaggerates the texture of the lizard's rough, sandpaper-like skin, covered in sharp, keratinized, conical spines. The edges and textures of each scale are clearly visible, and its intricate camouflage pattern of desert ochre, sandy yellow, and dark brown appears strikingly realistic under the blazing afternoon sun. It advances with a measured, stop-and-go rhythm, cautiously lifting a foot, pausing briefly in mid-air before firmly lowering it. Its small, obsidian eyes flick vigilantly, and its head moves slightly from side to side, as if searching for water. Subtle film grain dances across the frame, combined with a noticeable high-noise texture, imbuing this leisurely, focused scene with a powerful documentary quality, resembling a moment brimming with raw vitality captured from a top-tier nature documentary.", + "seed": 984, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_14/index_3660_seed_984_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": -0.12665553335100413, + "vlm_logprob": -0.12665553335100413, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.12665553335100413, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.3534717559814453 + ], + "opposite_logprob_tokens": [ + -1.2284717559814453 + ], + "logprob_reward": -0.3534717559814453, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3534717559814453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029442107304930687, + 0.12620526552200317, + 0.13015232980251312, + 0.05758515000343323, + 0.026205429807305336, + 0.08682803809642792, + 0.13318437337875366, + 0.01238724123686552, + 0.036388762295246124, + 0.36162132024765015 + ], + "credit_map": [ + 0.0028123536612838507, + 0.005101443268358707, + 0.0022694068029522896, + 0.002898797392845154, + 0.006965058855712414, + 0.0031489557586610317, + 0.00292544555850327, + 0.0033206467051059008, + 0.012056591920554638, + 0.02317420393228531, + 0.009988874197006226, + 0.017756301909685135, + 0.02196122147142887, + 0.012236006557941437, + 0.017315438017249107, + 0.0117166293784976, + 0.01609204337000847, + 0.01847848668694496, + 0.020166359841823578, + 0.018007677048444748, + 0.020357074216008186, + 0.016465894877910614, + 0.01115510892122984, + 0.00942969135940075, + 0.0075679258443415165, + 0.006130126304924488, + 0.009090820327401161, + 0.009321135468780994, + 0.006581255234777927, + 0.007568620145320892, + 0.0064541930332779884, + 0.004871073644608259, + 0.002757946029305458, + 0.003166372887790203, + 0.004068483132869005, + 0.0026171684730798006, + 0.0033651739358901978, + 0.004433723632246256, + 0.0032909144647419453, + 0.002505646087229252, + 0.007173972204327583, + 0.01159596536308527, + 0.009617290459573269, + 0.018124660477042198, + 0.011013747192919254, + 0.009682772681117058, + 0.008278738707304, + 0.011340892873704433, + 0.032938700169324875, + 0.008327733725309372, + 0.009154385887086391, + 0.008879344910383224, + 0.005261131562292576, + 0.009646441787481308, + 0.007859987206757069, + 0.05111663416028023, + 0.0023057227954268456, + 0.0014667571522295475, + 0.001481242710724473, + 0.0012360784457996488, + 0.001701947534456849, + 0.0016488030087202787, + 0.001158565515652299, + 0.001388123375363648, + 0.003817520337179303, + 0.005377215798944235, + 0.004440768621861935, + 0.007448278833180666, + 0.003508366644382477, + 0.00460586603730917, + 0.004496121779084206, + 0.0026946242433041334, + 0.023304423317313194, + 0.011602737940847874, + 0.023558666929602623, + 0.015647897496819496, + 0.025805875658988953, + 0.12421365082263947, + 0.1160208135843277, + 0.02146725170314312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.3534717559814453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.007657223381102085 + ], + "opposite_logprob_tokens": [ + -4.882657051086426 + ], + "logprob_reward": -0.007657223381102085, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007657223381102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15747500956058502, + 0.423127681016922, + 0.04477178677916527, + 0.043689366430044174, + 0.019594373181462288, + 0.04414913430809975, + 0.09441125392913818, + 0.02770547941327095, + 0.05568219721317291, + 0.08939367532730103 + ], + "credit_map": [ + 0.013734318315982819, + 0.01741071231663227, + 0.013401762582361698, + 0.02256067283451557, + 0.02372797019779682, + 0.019374536350369453, + 0.027365852147340775, + 0.01989918015897274, + 0.04323127120733261, + 0.08212772756814957, + 0.033034518361091614, + 0.05900569260120392, + 0.07201705873012543, + 0.04042171686887741, + 0.05567333474755287, + 0.03761634603142738, + 0.005023222416639328, + 0.006993745919317007, + 0.00662277452647686, + 0.005176618229597807, + 0.007814825512468815, + 0.005106552504003048, + 0.003898149821907282, + 0.004135897383093834, + 0.007266176864504814, + 0.0051755961030721664, + 0.0046609812416136265, + 0.006509581580758095, + 0.004989233799278736, + 0.004778096452355385, + 0.005293041467666626, + 0.005016659386456013, + 0.002674608025699854, + 0.0018490883521735668, + 0.004062075167894363, + 0.0023764732759445906, + 0.0018270360305905342, + 0.0022432843688875437, + 0.002429912332445383, + 0.0021318956278264523, + 0.004007410723716021, + 0.005088701844215393, + 0.005638736765831709, + 0.00864934828132391, + 0.005736848805099726, + 0.004821306094527245, + 0.004346359986811876, + 0.005860422272235155, + 0.020385392010211945, + 0.0054771569557487965, + 0.007420449983328581, + 0.007216308265924454, + 0.005064979661256075, + 0.00680233258754015, + 0.007269891444593668, + 0.03477473929524422, + 0.006011125165969133, + 0.003773970529437065, + 0.004433844238519669, + 0.002877749502658844, + 0.002423375379294157, + 0.002933982526883483, + 0.0023801675997674465, + 0.0028712644707411528, + 0.006492685526609421, + 0.010998629964888096, + 0.0064750248566269875, + 0.008182890713214874, + 0.005240555852651596, + 0.006917718332260847, + 0.006769650615751743, + 0.00460504274815321, + 0.005834662821143866, + 0.0038004526868462563, + 0.006677260156720877, + 0.0044973911717534065, + 0.007835241965949535, + 0.027805421501398087, + 0.02761291339993477, + 0.005330327432602644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.007657223381102085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.02329564467072487 + ], + "opposite_logprob_tokens": [ + -3.7732956409454346 + ], + "logprob_reward": -0.02329564467072487, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329564467072487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022013600915670395, + 0.6141800880432129, + 0.03370276466012001, + 0.03381279855966568, + 0.03947632387280464, + 0.023011766374111176, + 0.08187276870012283, + 0.007716088555753231, + 0.022351091727614403, + 0.12186272442340851 + ], + "credit_map": [ + 0.002664078725501895, + 0.0029270383529365063, + 0.0028006918728351593, + 0.002570749493315816, + 0.0027868992183357477, + 0.002305941190570593, + 0.0033395783975720406, + 0.0026186220347881317, + 0.06951358914375305, + 0.11579341441392899, + 0.0504407063126564, + 0.08090364187955856, + 0.09140852093696594, + 0.05959199741482735, + 0.03945573419332504, + 0.10707251727581024, + 0.0037714671343564987, + 0.00374031369574368, + 0.006863540504127741, + 0.004476552363485098, + 0.004614342004060745, + 0.00447539146989584, + 0.0034287692978978157, + 0.002332386327907443, + 0.005151159130036831, + 0.003327742451801896, + 0.004757638089358807, + 0.005807808134704828, + 0.0033101909793913364, + 0.00441316282376647, + 0.003997719381004572, + 0.0030473796650767326, + 0.0041906461119651794, + 0.003739273874089122, + 0.005912691354751587, + 0.0034454769920557737, + 0.006096964702010155, + 0.007507896516472101, + 0.0046800971031188965, + 0.003903280710801482, + 0.0022933140862733126, + 0.0030258509796112776, + 0.0024942224845290184, + 0.002580045722424984, + 0.0033731081057339907, + 0.0039047691971063614, + 0.0025564173702150583, + 0.0027840379625558853, + 0.020153310149908066, + 0.004840970505028963, + 0.004086843691766262, + 0.0038231443613767624, + 0.004446139559149742, + 0.0055242604576051235, + 0.005472569726407528, + 0.033525533974170685, + 0.0018491065129637718, + 0.0007401249604299664, + 0.0009356248774565756, + 0.0007035158341750503, + 0.0009731940808705986, + 0.0010844918433576822, + 0.0008532584761269391, + 0.0005767722032032907, + 0.0019531482830643654, + 0.0025963177904486656, + 0.003099781461060047, + 0.005578884854912758, + 0.001874225097708404, + 0.0037513677962124348, + 0.0016604847041890025, + 0.0018368810415267944, + 0.007547252811491489, + 0.0032052109017968178, + 0.00777271157130599, + 0.004576551727950573, + 0.006454382557421923, + 0.04390328750014305, + 0.04080606624484062, + 0.007597256917506456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.02329564467072487 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.023299604654312134 + ], + "opposite_logprob_tokens": [ + -3.7732996940612793 + ], + "logprob_reward": -0.023299604654312134, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299604654312134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06849800795316696, + 0.22690331935882568, + 0.0193481408059597, + 0.03536713495850563, + 0.061459604650735855, + 0.06453153491020203, + 0.1141742542386055, + 0.017288047820329666, + 0.1776607483625412, + 0.21476921439170837 + ], + "credit_map": [ + 0.007712493184953928, + 0.009333728812634945, + 0.006581189576536417, + 0.008739364333450794, + 0.008694430813193321, + 0.008231425657868385, + 0.00878226663917303, + 0.010423109866678715, + 0.023469287902116776, + 0.03544089198112488, + 0.016855129972100258, + 0.029194576665759087, + 0.029336625710129738, + 0.021758567541837692, + 0.019259802997112274, + 0.05158842355012894, + 0.002651916118338704, + 0.00286004482768476, + 0.0034601474180817604, + 0.00195861654356122, + 0.002749726176261902, + 0.002217644127085805, + 0.0014484518906101584, + 0.002001595450565219, + 0.005388940218836069, + 0.0031437703873962164, + 0.004888149909675121, + 0.00614176457747817, + 0.003017978509888053, + 0.004862846806645393, + 0.004312301054596901, + 0.003611381631344557, + 0.005756078753620386, + 0.0036486140452325344, + 0.010959338396787643, + 0.004201710689812899, + 0.009412875398993492, + 0.011691580526530743, + 0.010348636656999588, + 0.005440768785774708, + 0.00557850394397974, + 0.010302369482815266, + 0.004854622296988964, + 0.008182837627828121, + 0.007634220644831657, + 0.00950698647648096, + 0.00843952875584364, + 0.01003246195614338, + 0.030928025022149086, + 0.004693618509918451, + 0.0057937889359891415, + 0.005261019803583622, + 0.005326718557626009, + 0.007964118383824825, + 0.007072689477354288, + 0.047134269028902054, + 0.002741953358054161, + 0.002287423936650157, + 0.0022825158666819334, + 0.0019804267212748528, + 0.0023517815861850977, + 0.001702577224932611, + 0.0016572520835325122, + 0.0022841182071715593, + 0.015537422150373459, + 0.02252851240336895, + 0.02216501720249653, + 0.049000754952430725, + 0.013317347504198551, + 0.02311595343053341, + 0.018223188817501068, + 0.013772543519735336, + 0.013093598186969757, + 0.005676427856087685, + 0.013241689652204514, + 0.007972782477736473, + 0.011626866646111012, + 0.07762134820222855, + 0.07181481271982193, + 0.013721682131290436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.023299604654312134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.22555343806743622 + ], + "opposite_logprob_tokens": [ + -1.6005533933639526 + ], + "logprob_reward": -0.22555343806743622, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22555343806743622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05547095462679863, + 0.5518021583557129, + 0.02730633318424225, + 0.033582158386707306, + 0.04454485699534416, + 0.03987222909927368, + 0.062179166823625565, + 0.030009273439645767, + 0.040578313171863556, + 0.11465451866388321 + ], + "credit_map": [ + 0.0063519724644720554, + 0.011033128947019577, + 0.005670057609677315, + 0.00491414126008749, + 0.010943949222564697, + 0.0070527647621929646, + 0.005141241010278463, + 0.004363703075796366, + 0.06087635084986687, + 0.10567037761211395, + 0.04238250479102135, + 0.07667836546897888, + 0.0870337188243866, + 0.05312150716781616, + 0.03921029716730118, + 0.08682902157306671, + 0.003581342985853553, + 0.00510291988030076, + 0.003094281069934368, + 0.0032468705903738737, + 0.0043557207100093365, + 0.002792135113850236, + 0.0028075287118554115, + 0.0023255350533872843, + 0.004873943980783224, + 0.003384131472557783, + 0.00361529435031116, + 0.0038516439963132143, + 0.003288976615294814, + 0.005425958428531885, + 0.004894821438938379, + 0.00424738647416234, + 0.004454835783690214, + 0.0032825907692313194, + 0.009306230582296848, + 0.003871010383591056, + 0.0063444990664720535, + 0.007706830278038979, + 0.005677817855030298, + 0.0039010460022836924, + 0.0028101634234189987, + 0.004087495151907206, + 0.004190005362033844, + 0.006759403273463249, + 0.004205901175737381, + 0.00599872088059783, + 0.0057790810242295265, + 0.0060414583422243595, + 0.011285305954515934, + 0.0035090195015072823, + 0.00439805630594492, + 0.004279287066310644, + 0.00327430316247046, + 0.005761872977018356, + 0.004215946886688471, + 0.025455374270677567, + 0.00677039148285985, + 0.004199178423732519, + 0.004686418455094099, + 0.0024789487943053246, + 0.0036063888110220432, + 0.002837660489603877, + 0.00238046795129776, + 0.0030498190317302942, + 0.004388666711747646, + 0.005213391967117786, + 0.005632416345179081, + 0.008714854717254639, + 0.0030937260016798973, + 0.004103057086467743, + 0.0060449205338954926, + 0.003387278877198696, + 0.007740594446659088, + 0.0034011106472462416, + 0.007514205761253834, + 0.004941852763295174, + 0.006846501957625151, + 0.03927820920944214, + 0.0376901850104332, + 0.007241853978484869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.22555343806743622 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, cinematic close-up, with a slow, almost imperceptible pan to the right, closely follows an Australian Gila monster as it slowly searches for water in the dry, cracked red earth. The sharpened image exaggerates the texture of the lizard's rough, sandpaper-like skin, covered in sharp, keratinized, conical spines. The edges and textures of each scale are clearly visible, and its intricate camouflage pattern of desert ochre, sandy yellow, and dark brown appears strikingly realistic under the blazing afternoon sun. It advances with a measured, stop-and-go rhythm, cautiously lifting a foot, pausing briefly in mid-air before firmly lowering it. Its small, obsidian eyes flick vigilantly, and its head moves slightly from side to side, as if searching for water. Subtle film grain dances across the frame, combined with a noticeable high-noise texture, imbuing this leisurely, focused scene with a powerful documentary quality, resembling a moment brimming with raw vitality captured from a top-tier nature documentary.", + "seed": 985, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_14/index_3660_seed_985_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": -0.1640784304589033, + "vlm_logprob": -0.1640784304589033, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.1640784304589033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.39181193709373474 + ], + "opposite_logprob_tokens": [ + -1.1418119668960571 + ], + "logprob_reward": -0.39181193709373474, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.39181193709373474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.485811710357666, + 0.0794445127248764, + 0.05588192120194435, + 0.044992439448833466, + 0.07082370668649673, + 0.02954195812344551, + 0.06609288603067398, + 0.03205447643995285, + 0.09055636078119278, + 0.04480011388659477 + ], + "credit_map": [ + 0.08046824485063553, + 0.06048191338777542, + 0.08825173228979111, + 0.06923342496156693, + 0.05375320091843605, + 0.045898690819740295, + 0.04808609560132027, + 0.03963836282491684, + 0.007526399102061987, + 0.012787692248821259, + 0.00980730913579464, + 0.010060017928481102, + 0.011850579641759396, + 0.010044571943581104, + 0.006876231636852026, + 0.010491698049008846, + 0.0049242982640862465, + 0.003800048725679517, + 0.008437277749180794, + 0.00630074692890048, + 0.0069790068082511425, + 0.012993917800486088, + 0.0055618928745388985, + 0.0068847183138132095, + 0.007429652381688356, + 0.004652060102671385, + 0.004591998644173145, + 0.006466171704232693, + 0.0056494781747460365, + 0.0034520074259489775, + 0.00825635064393282, + 0.004494712688028812, + 0.01104286964982748, + 0.009683242067694664, + 0.00903795100748539, + 0.007854904048144817, + 0.006247215438634157, + 0.006195052061229944, + 0.01376291736960411, + 0.006999541539698839, + 0.004104812163859606, + 0.005037991795688868, + 0.002271168865263462, + 0.003911146428436041, + 0.00434507941827178, + 0.0026747744996100664, + 0.003605849342420697, + 0.003591128159314394, + 0.0061593796126544476, + 0.006495833862572908, + 0.0064853886142373085, + 0.010395360179245472, + 0.00872842874377966, + 0.006431862711906433, + 0.013207495212554932, + 0.008189124055206776, + 0.007759748958051205, + 0.004759219940751791, + 0.0029879019130021334, + 0.0038020857609808445, + 0.0037799281999468803, + 0.0028290662448853254, + 0.0033250588458031416, + 0.002811458660289645, + 0.014540765434503555, + 0.014183719642460346, + 0.012591968290507793, + 0.013904868625104427, + 0.009567773900926113, + 0.0072765592485666275, + 0.011514549143612385, + 0.006976142060011625, + 0.005182548891752958, + 0.006699798163026571, + 0.004661265294998884, + 0.0033886374440044165, + 0.006240657065063715, + 0.007045418955385685, + 0.007935958914458752, + 0.0036458231043070555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.39181193709373474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.016102386638522148 + ], + "opposite_logprob_tokens": [ + -4.141102313995361 + ], + "logprob_reward": -0.016102386638522148, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016102386638522148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3781554102897644, + 0.05855950340628624, + 0.06185324117541313, + 0.07809772342443466, + 0.07914265990257263, + 0.03298114240169525, + 0.091123066842556, + 0.11625993996858597, + 0.025428803637623787, + 0.07839852571487427 + ], + "credit_map": [ + 0.06087536737322807, + 0.048094335943460464, + 0.06624140590429306, + 0.05492516979575157, + 0.03611556068062782, + 0.0326969251036644, + 0.046086959540843964, + 0.03311968967318535, + 0.00730028236284852, + 0.005168232135474682, + 0.006681119557470083, + 0.012284720316529274, + 0.004996687639504671, + 0.005133906379342079, + 0.007142331451177597, + 0.009852222166955471, + 0.004498151130974293, + 0.0037405688781291246, + 0.007378743030130863, + 0.006809850223362446, + 0.007146084681153297, + 0.017873408272862434, + 0.0054695699363946915, + 0.008936863392591476, + 0.013777262531220913, + 0.007292240858078003, + 0.00842620525509119, + 0.009072495624423027, + 0.009994530119001865, + 0.006408398970961571, + 0.015187613666057587, + 0.00793896708637476, + 0.003970992285758257, + 0.0044182115234434605, + 0.004960157442837954, + 0.01011215802282095, + 0.006668912712484598, + 0.007867592386901379, + 0.0277993306517601, + 0.013345298357307911, + 0.0029012192972004414, + 0.003610439831390977, + 0.0026114429347217083, + 0.005692039616405964, + 0.0029017215128988028, + 0.004164077807217836, + 0.004347122740000486, + 0.006753076333552599, + 0.008115341886878014, + 0.008527131751179695, + 0.015473763458430767, + 0.012606292963027954, + 0.00731967156752944, + 0.01294107735157013, + 0.015590706840157509, + 0.010549081489443779, + 0.021323062479496002, + 0.01586245372891426, + 0.011843044310808182, + 0.019064296036958694, + 0.010678150691092014, + 0.012759042903780937, + 0.014543299563229084, + 0.010186582803726196, + 0.00355194928124547, + 0.003098265500739217, + 0.003481480060145259, + 0.0032516815699636936, + 0.003131561679765582, + 0.002326536923646927, + 0.00263228896073997, + 0.0039550368674099445, + 0.014449357986450195, + 0.009538847953081131, + 0.0126662477850914, + 0.008691372349858284, + 0.00686790281906724, + 0.008506321348249912, + 0.009639259427785873, + 0.008039211854338646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.016102386638522148 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.0181988924741745 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.0181988924741745, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181988924741745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07364710420370102, + 0.128277987241745, + 0.06559256464242935, + 0.05835605040192604, + 0.06677427142858505, + 0.060466207563877106, + 0.14267925918102264, + 0.03196573257446289, + 0.07338262349367142, + 0.2988581359386444 + ], + "credit_map": [ + 0.013273695483803749, + 0.01073931809514761, + 0.014618461951613426, + 0.009081847965717316, + 0.00737401656806469, + 0.007587403059005737, + 0.0045939539559185505, + 0.006378402467817068, + 0.01259223185479641, + 0.019909890368580818, + 0.013572020456194878, + 0.011865007691085339, + 0.02031058445572853, + 0.022086817771196365, + 0.011471729725599289, + 0.016469700261950493, + 0.005046308971941471, + 0.006384653504937887, + 0.009017580188810825, + 0.011078111827373505, + 0.007740195840597153, + 0.010093779303133488, + 0.007990642450749874, + 0.008241291157901287, + 0.009034904651343822, + 0.00493200309574604, + 0.009173087775707245, + 0.007425387389957905, + 0.006855108309537172, + 0.005192354321479797, + 0.00957927294075489, + 0.006163928657770157, + 0.0076949880458414555, + 0.006609758362174034, + 0.0078087253496050835, + 0.009177862666547298, + 0.0056922296062111855, + 0.005707503296434879, + 0.0149771086871624, + 0.009106095880270004, + 0.010190392844378948, + 0.00827097613364458, + 0.007276667281985283, + 0.007920833304524422, + 0.005818687845021486, + 0.005460753571242094, + 0.009115936234593391, + 0.006411961745470762, + 0.012093126773834229, + 0.020214684307575226, + 0.018308885395526886, + 0.0172327421605587, + 0.01864784210920334, + 0.014916334301233292, + 0.025175923481583595, + 0.01608973555266857, + 0.007642797194421291, + 0.003780853468924761, + 0.0025246436707675457, + 0.004616330377757549, + 0.003186493180692196, + 0.0030111796222627163, + 0.00441711163148284, + 0.0027863248251378536, + 0.012236013077199459, + 0.011945678852498531, + 0.012734737247228622, + 0.009533964097499847, + 0.007147032301872969, + 0.005001645535230637, + 0.008253982290625572, + 0.006529574282467365, + 0.00932220183312893, + 0.009677272289991379, + 0.008827578276395798, + 0.012132089585065842, + 0.011618070304393768, + 0.0874839574098587, + 0.12726594507694244, + 0.032531045377254486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.0181988924741745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.11296396702528 + ], + "opposite_logprob_tokens": [ + -2.237963914871216 + ], + "logprob_reward": -0.11296396702528, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296396702528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18521352112293243, + 0.09178624302148819, + 0.07853075861930847, + 0.16224567592144012, + 0.05362351983785629, + 0.028658287599682808, + 0.06518305838108063, + 0.08228424191474915, + 0.05157146975398064, + 0.20090316236019135 + ], + "credit_map": [ + 0.030941424891352654, + 0.022187264636158943, + 0.0355457179248333, + 0.026280680671334267, + 0.018419481813907623, + 0.01592579111456871, + 0.019880473613739014, + 0.016032690182328224, + 0.011383085511624813, + 0.012272861786186695, + 0.008273415267467499, + 0.010236328467726707, + 0.012623799033463001, + 0.01447284035384655, + 0.008372082374989986, + 0.01415182463824749, + 0.008810273371636868, + 0.009361078031361103, + 0.008876503445208073, + 0.01282460056245327, + 0.01030636578798294, + 0.01074189506471157, + 0.008658327162265778, + 0.008951717987656593, + 0.030651919543743134, + 0.014909257180988789, + 0.016851188614964485, + 0.016328632831573486, + 0.0278693288564682, + 0.012537351809442043, + 0.025997284799814224, + 0.01710071600973606, + 0.0037955110892653465, + 0.004240958020091057, + 0.005863932427018881, + 0.006420762278139591, + 0.004984179977327585, + 0.005482336971908808, + 0.014959909953176975, + 0.007875930517911911, + 0.0038290699012577534, + 0.003972102887928486, + 0.00202074833214283, + 0.0034957416355609894, + 0.0027275464963167906, + 0.0037571967113763094, + 0.0045650121755898, + 0.004290867131203413, + 0.0053436667658388615, + 0.005753738339990377, + 0.007876608520746231, + 0.01117704901844263, + 0.0072125038132071495, + 0.006988250184804201, + 0.013226808980107307, + 0.007604434620589018, + 0.014437628909945488, + 0.012514668516814709, + 0.006743907928466797, + 0.013105127960443497, + 0.00788041204214096, + 0.007899925112724304, + 0.011174862273037434, + 0.00852771196514368, + 0.00839949306100607, + 0.00677204504609108, + 0.006975376047194004, + 0.007082055322825909, + 0.005188979208469391, + 0.005725398194044828, + 0.005981938447803259, + 0.005446184892207384, + 0.0219433531165123, + 0.011999244801700115, + 0.017658831551671028, + 0.011534703895449638, + 0.009096701629459858, + 0.05111110210418701, + 0.061904117465019226, + 0.015655113384127617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.11296396702528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 Australian Gila monster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment dry, cracked red earth under a blazing afternoon sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lizard move with a measured, stop-and-go rhythm, lifting a foot, pausing, and lowering it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image characterized by high-noise texture and subtle film grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, almost imperceptible pan to the right following 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": [ + -0.2813149690628052 + ], + "opposite_logprob_tokens": [ + -1.4063149690628052 + ], + "logprob_reward": -0.2813149690628052, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813149690628052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08974415808916092, + 0.06502778828144073, + 0.11976771056652069, + 0.05127957835793495, + 0.11117898672819138, + 0.08096088469028473, + 0.15984846651554108, + 0.06676620990037918, + 0.04123074188828468, + 0.21419544517993927 + ], + "credit_map": [ + 0.016053127124905586, + 0.012293096631765366, + 0.017670873552560806, + 0.011430975049734116, + 0.007930554449558258, + 0.007401265669614077, + 0.009153741411864758, + 0.00781052652746439, + 0.007519765291363001, + 0.0071134124882519245, + 0.007330581545829773, + 0.009542533196508884, + 0.007281369995325804, + 0.008726640604436398, + 0.009340868331491947, + 0.008172622881829739, + 0.015575810335576534, + 0.013350351713597775, + 0.014137404039502144, + 0.026976043358445168, + 0.012955239042639732, + 0.010467330925166607, + 0.01230509951710701, + 0.01400044560432434, + 0.006714345887303352, + 0.006834247149527073, + 0.004424701910465956, + 0.009412831626832485, + 0.005518137011677027, + 0.004574500024318695, + 0.008429177105426788, + 0.0053716436959803104, + 0.010065185837447643, + 0.008761605247855186, + 0.008705715648829937, + 0.01304212212562561, + 0.010355434380471706, + 0.01164928637444973, + 0.03128170594573021, + 0.017317939549684525, + 0.013498665764927864, + 0.012017585337162018, + 0.006806482560932636, + 0.009565964341163635, + 0.009851445443928242, + 0.009818101301789284, + 0.01020490936934948, + 0.009197738021612167, + 0.01664738543331623, + 0.015852665528655052, + 0.015791499987244606, + 0.02132372558116913, + 0.025904985144734383, + 0.016889818012714386, + 0.02916674129664898, + 0.018271656706929207, + 0.01138442475348711, + 0.009767035022377968, + 0.007083822973072529, + 0.013891384936869144, + 0.006053335499018431, + 0.005771327763795853, + 0.006492507178336382, + 0.006322378292679787, + 0.0073204743675887585, + 0.005422887858003378, + 0.005877560470253229, + 0.005559013690799475, + 0.003961970563977957, + 0.004120274446904659, + 0.004565711133182049, + 0.004402851220220327, + 0.022532880306243896, + 0.01800098828971386, + 0.021563850343227386, + 0.012350894510746002, + 0.01617187261581421, + 0.0568494088947773, + 0.031268633902072906, + 0.03545691817998886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.2813149690628052 + } + ], + "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/0000110/group_15/index_tensor([5129, 5129])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_15/index_tensor([5129, 5129])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6cc5ba4b156b0b75e661b0ceb1e1c990cebbb5a9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_15/index_tensor([5129, 5129])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely close-up, static macro shot captures a pool of smooth, dark syrup as it slowly solidifies on a cool, matte gray surface, the entire scene rendered in dramatic, high-contrast black and white. A harsh key light sweeps across the frame from a low side angle, casting deep, velvety shadows and creating brilliant, mirrored highlights across the syrup's initially viscous, mirrored surface. As the fluid edges of the liquid begin to solidify, the transformation begins, and the once perfectly smooth areas begin to fracture from within. Intricate crystalline structures emerge and branch like frost patterns, their sharp facets catching the harsh light, causing them to shimmer brilliantly white against the deep black of the solidifying substance. The camera remains perfectly still, allowing the viewer to witness the mesmerizing, almost geological process of the organic liquid's inexorable hardening, its texture transforming from smooth and glossy to a complex, shimmering, fractured solid, creating an intense, abstract atmosphere.", + "seed": 11901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_15/index_5129_seed_11901_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": -0.3836479078279808, + "vlm_logprob": -0.3836479078279808, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.3836479078279808, + "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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.2042442113161087 + ], + "opposite_logprob_tokens": [ + -1.7042442560195923 + ], + "logprob_reward": -0.2042442113161087, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2042442113161087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1086215078830719, + 0.36223357915878296, + 0.058826204389333725, + 0.05059073865413666, + 0.18599627912044525, + 0.03642493858933449, + 0.06197907403111458, + 0.042794033885002136, + 0.0635981559753418, + 0.02893545851111412 + ], + "credit_map": [ + 0.01212933287024498, + 0.013445965014398098, + 0.008642682805657387, + 0.01821153797209263, + 0.0106572937220335, + 0.01370304450392723, + 0.021038705483078957, + 0.010792943648993969, + 0.040793124586343765, + 0.023621011525392532, + 0.02360544539988041, + 0.06470219790935516, + 0.017651071771979332, + 0.1434302031993866, + 0.024507805705070496, + 0.023922711610794067, + 0.006750206928700209, + 0.008885602466762066, + 0.004823741968721151, + 0.009954598732292652, + 0.0059822481125593185, + 0.005522481631487608, + 0.00806565210223198, + 0.008841674774885178, + 0.006091577000916004, + 0.005795094650238752, + 0.005304412450641394, + 0.006009003147482872, + 0.00768189737573266, + 0.004493841901421547, + 0.008972225710749626, + 0.006242692470550537, + 0.02118062414228916, + 0.02500796690583229, + 0.02466755546629429, + 0.03378979489207268, + 0.020268213003873825, + 0.017420321702957153, + 0.023159803822636604, + 0.020502004772424698, + 0.006005135830491781, + 0.0036327068228274584, + 0.004765199963003397, + 0.00398977380245924, + 0.005142402369529009, + 0.004479138646274805, + 0.003795316908508539, + 0.004615264944732189, + 0.00536034069955349, + 0.0069559914991259575, + 0.0078920042142272, + 0.010276698507368565, + 0.00568490382283926, + 0.012789394706487656, + 0.0055087232030928135, + 0.0075110141187906265, + 0.0035306434147059917, + 0.006468123756349087, + 0.00728248618543148, + 0.007128136232495308, + 0.005241876933723688, + 0.004215217661112547, + 0.0041426257230341434, + 0.004784923978149891, + 0.007417589426040649, + 0.008221929892897606, + 0.010422585532069206, + 0.011723061092197895, + 0.0061879754066467285, + 0.007051580585539341, + 0.00569122564047575, + 0.006882215850055218, + 0.0029384223744273186, + 0.0031896138098090887, + 0.003467240370810032, + 0.0028598487842828035, + 0.0038062920793890953, + 0.005823798943310976, + 0.0026169628836214542, + 0.004233278799802065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.2042442113161087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.011117065325379372 + ], + "opposite_logprob_tokens": [ + -4.511116981506348 + ], + "logprob_reward": -0.011117065325379372, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011117065325379372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2416052520275116, + 0.2870928645133972, + 0.14616771042346954, + 0.0687556192278862, + 0.06901659071445465, + 0.02513330988585949, + 0.023785369470715523, + 0.015339345671236515, + 0.02621331624686718, + 0.0968906506896019 + ], + "credit_map": [ + 0.0252930149435997, + 0.024794764816761017, + 0.016103368252515793, + 0.04435352981090546, + 0.021674850955605507, + 0.04509066045284271, + 0.04030677676200867, + 0.0239882692694664, + 0.03253712132573128, + 0.022763311862945557, + 0.02382802590727806, + 0.050078537315130234, + 0.019415393471717834, + 0.10097573697566986, + 0.01884082704782486, + 0.01865391805768013, + 0.01580333709716797, + 0.0193875040858984, + 0.013401017524302006, + 0.026772474870085716, + 0.012221441604197025, + 0.013606547378003597, + 0.020299367606639862, + 0.024676010012626648, + 0.007222714833915234, + 0.006421754602342844, + 0.006205150857567787, + 0.005823837127536535, + 0.014435872435569763, + 0.010472681373357773, + 0.009205542504787445, + 0.008968060836195946, + 0.007952076382935047, + 0.006087685003876686, + 0.008930306881666183, + 0.005867167375981808, + 0.01529427245259285, + 0.00925289373844862, + 0.009594028815627098, + 0.006038154475390911, + 0.0035568352323025465, + 0.0028027992229908705, + 0.003776231547817588, + 0.003303767181932926, + 0.003869791515171528, + 0.0024478991981595755, + 0.002495428314432502, + 0.0028805548790842295, + 0.001967702293768525, + 0.0034750550985336304, + 0.003001127392053604, + 0.004573946353048086, + 0.0026911681052297354, + 0.0026005739346146584, + 0.0024684357922524214, + 0.003007358405739069, + 0.0016179517842829227, + 0.0016150623559951782, + 0.0015779148088768125, + 0.001985497772693634, + 0.002047107554972172, + 0.003213617717847228, + 0.0013395929709076881, + 0.001942600472830236, + 0.002288609044626355, + 0.00390621367841959, + 0.003910824190825224, + 0.004118535667657852, + 0.002523409202694893, + 0.003942475188523531, + 0.0023788802791386843, + 0.003144368762150407, + 0.010420413687825203, + 0.010416622273623943, + 0.011953677982091904, + 0.009361824952065945, + 0.010291227139532566, + 0.020916946232318878, + 0.009202554821968079, + 0.014327378943562508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.011117065325379372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.10036207735538483 + ], + "opposite_logprob_tokens": [ + -2.3503620624542236 + ], + "logprob_reward": -0.10036207735538483, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036207735538483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048964545130729675, + 0.5790202021598816, + 0.10266026109457016, + 0.03276296332478523, + 0.05043010786175728, + 0.07706207782030106, + 0.06605717539787292, + 0.006804945878684521, + 0.018445536494255066, + 0.017792288213968277 + ], + "credit_map": [ + 0.0062698195688426495, + 0.007543645333498716, + 0.004059670027345419, + 0.0077248262241482735, + 0.005052896682173014, + 0.006454042624682188, + 0.006462904158979654, + 0.005396741442382336, + 0.06696407496929169, + 0.03667066618800163, + 0.037727124989032745, + 0.10230225324630737, + 0.028175929561257362, + 0.23466017842292786, + 0.035719383507966995, + 0.036800555884838104, + 0.010993932373821735, + 0.014021120965480804, + 0.008744773454964161, + 0.01739085651934147, + 0.009696358814835548, + 0.009611452929675579, + 0.014678244479000568, + 0.017523519694805145, + 0.0038600279949605465, + 0.002994891256093979, + 0.0021116817370057106, + 0.002173675922676921, + 0.003655784297734499, + 0.0030237766914069653, + 0.005635614041239023, + 0.009307513013482094, + 0.0059524220414459705, + 0.007118596229702234, + 0.005584423895925283, + 0.009327355772256851, + 0.006653099786490202, + 0.004359993152320385, + 0.006706212647259235, + 0.004728003405034542, + 0.01639040745794773, + 0.007005604915320873, + 0.010156537406146526, + 0.007985922507941723, + 0.007963796146214008, + 0.010159078985452652, + 0.008517294190824032, + 0.008883433416485786, + 0.006382028106600046, + 0.008496319875121117, + 0.007649568375200033, + 0.012242445722222328, + 0.008589041419327259, + 0.009280207566916943, + 0.00524559710174799, + 0.008171964436769485, + 0.0007150642923079431, + 0.0005935137160122395, + 0.0008046259172260761, + 0.0006419704877771437, + 0.0008060730178840458, + 0.0016969064017757773, + 0.0007349130464717746, + 0.000811878708191216, + 0.0015237004263326526, + 0.003079085610806942, + 0.0019413059344515204, + 0.0019783100578933954, + 0.0015582533087581396, + 0.00392654025927186, + 0.0018552833935245872, + 0.002583058550953865, + 0.0021102859172970057, + 0.0020670797675848007, + 0.0017688608495518565, + 0.0016509044216945767, + 0.001986858667805791, + 0.003865055274218321, + 0.0015899845166131854, + 0.0027532575186342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.10036207735538483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.0019469369435682893 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.0019469369435682893, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019469369435682893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12388306111097336, + 0.28387415409088135, + 0.09182266145944595, + 0.05114656314253807, + 0.14990100264549255, + 0.03337942808866501, + 0.05264170467853546, + 0.05161113664507866, + 0.09706719219684601, + 0.06467314064502716 + ], + "credit_map": [ + 0.012800225056707859, + 0.013038485310971737, + 0.009722596034407616, + 0.022822003811597824, + 0.012460323050618172, + 0.01899963803589344, + 0.021821698173880577, + 0.012218094430863857, + 0.02920789085328579, + 0.022247958928346634, + 0.020807746797800064, + 0.052436742931604385, + 0.01820472627878189, + 0.09459783136844635, + 0.024419518187642097, + 0.02195173129439354, + 0.01085720956325531, + 0.01239997148513794, + 0.008121415972709656, + 0.015127511695027351, + 0.008206047117710114, + 0.008372209966182709, + 0.011945139616727829, + 0.016793156042695045, + 0.00566283892840147, + 0.0045766145922243595, + 0.004667780362069607, + 0.0042348820716142654, + 0.008623175323009491, + 0.006163721438497305, + 0.007715753745287657, + 0.009501799941062927, + 0.018771568313241005, + 0.02251191809773445, + 0.02097342535853386, + 0.029201794415712357, + 0.012606077827513218, + 0.01298652496188879, + 0.01829182542860508, + 0.014557873830199242, + 0.004379581660032272, + 0.0027567774523049593, + 0.002533865161240101, + 0.003762554842978716, + 0.006487822160124779, + 0.005896780174225569, + 0.0034280598629266024, + 0.004133987706154585, + 0.004207531921565533, + 0.005091201979666948, + 0.004774296656250954, + 0.011129319667816162, + 0.005378643982112408, + 0.009848630987107754, + 0.005860192235559225, + 0.0063518923707306385, + 0.005558093078434467, + 0.0047783623449504375, + 0.005647239275276661, + 0.008629240095615387, + 0.00616943696513772, + 0.00960563775151968, + 0.0050759026780724525, + 0.006147229578346014, + 0.010587628930807114, + 0.010488202795386314, + 0.015875203534960747, + 0.018999917432665825, + 0.00965623464435339, + 0.014274419285356998, + 0.00792560912668705, + 0.009259972721338272, + 0.006323236972093582, + 0.005548164248466492, + 0.007146414369344711, + 0.005995324347168207, + 0.00748577481135726, + 0.015783844515681267, + 0.006325713824480772, + 0.010064670816063881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0019469369435682893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.600569248199463 + ], + "opposite_logprob_tokens": [ + -0.2255692332983017 + ], + "logprob_reward": -1.600569248199463, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.600569248199463, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1207406148314476, + 0.19728049635887146, + 0.062096867710351944, + 0.10108017921447754, + 0.2038254290819168, + 0.07823709398508072, + 0.055820658802986145, + 0.030234279111027718, + 0.08483581990003586, + 0.06584858149290085 + ], + "credit_map": [ + 0.02013503760099411, + 0.012916837818920612, + 0.009394095279276371, + 0.016419196501374245, + 0.009825888089835644, + 0.0246588122099638, + 0.014783457852900028, + 0.012607290409505367, + 0.022345980629324913, + 0.014105619862675667, + 0.013798863627016544, + 0.030329490080475807, + 0.011222354136407375, + 0.07677599787712097, + 0.015636727213859558, + 0.013065462000668049, + 0.006308641284704208, + 0.008895265869796276, + 0.005965204909443855, + 0.010510438121855259, + 0.0058218645863235, + 0.0075834766030311584, + 0.00865916907787323, + 0.00835280492901802, + 0.011426728218793869, + 0.01266478467732668, + 0.008687958121299744, + 0.009672759100794792, + 0.012031969614326954, + 0.007528979331254959, + 0.020023537799715996, + 0.019043464213609695, + 0.021035902202129364, + 0.025062309578061104, + 0.025105580687522888, + 0.03139037266373634, + 0.02915981225669384, + 0.018925698474049568, + 0.027837133035063744, + 0.025308620184659958, + 0.010905366390943527, + 0.00830184668302536, + 0.010704250074923038, + 0.01402959879487753, + 0.007109368219971657, + 0.008452842012047768, + 0.008598553948104382, + 0.010135267861187458, + 0.005043833050876856, + 0.007695913314819336, + 0.006257543805986643, + 0.0063103497959673405, + 0.008088288828730583, + 0.009050103835761547, + 0.0058654979802668095, + 0.007509129587560892, + 0.0030895175877958536, + 0.0029448699206113815, + 0.003173672826960683, + 0.0033279319759458303, + 0.003889666171744466, + 0.005921356845647097, + 0.003708211239427328, + 0.00417905580252409, + 0.00900411605834961, + 0.013269658200442791, + 0.007498292718082666, + 0.01013864018023014, + 0.009013360366225243, + 0.01775539293885231, + 0.007200912106782198, + 0.010955451987683773, + 0.00544142397120595, + 0.00477976119145751, + 0.00891807023435831, + 0.006364183500409126, + 0.0070518250577151775, + 0.016384506598114967, + 0.006720845587551594, + 0.010187970474362373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -1.600569248199463 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extremely close-up, static macro shot captures a pool of smooth, dark syrup as it slowly solidifies on a cool, matte gray surface, the entire scene rendered in dramatic, high-contrast black and white. A harsh key light sweeps across the frame from a low side angle, casting deep, velvety shadows and creating brilliant, mirrored highlights across the syrup's initially viscous, mirrored surface. As the fluid edges of the liquid begin to solidify, the transformation begins, and the once perfectly smooth areas begin to fracture from within. Intricate crystalline structures emerge and branch like frost patterns, their sharp facets catching the harsh light, causing them to shimmer brilliantly white against the deep black of the solidifying substance. The camera remains perfectly still, allowing the viewer to witness the mesmerizing, almost geological process of the organic liquid's inexorable hardening, its texture transforming from smooth and glossy to a complex, shimmering, fractured solid, creating an intense, abstract atmosphere.", + "seed": 11902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_15/index_5129_seed_11902_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": -0.25121095853392034, + "vlm_logprob": -0.25121095853392034, + "vlm_yes_no_margin": 3.374999988079071, + "vlm_token_credit": -0.25121095853392034, + "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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.04278075322508812 + ], + "opposite_logprob_tokens": [ + -3.292780637741089 + ], + "logprob_reward": -0.04278075322508812, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04278075322508812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15896813571453094, + 0.24172726273536682, + 0.06483905762434006, + 0.05287620425224304, + 0.2664256691932678, + 0.06154606491327286, + 0.05740394815802574, + 0.016174357384443283, + 0.02181490883231163, + 0.05822443589568138 + ], + "credit_map": [ + 0.02025698497891426, + 0.029623880982398987, + 0.016245510429143906, + 0.01709788106381893, + 0.018245134502649307, + 0.019784267991781235, + 0.017416495829820633, + 0.020297972485423088, + 0.03380730003118515, + 0.022746002301573753, + 0.035905349999666214, + 0.02671917714178562, + 0.027170775458216667, + 0.030211547389626503, + 0.02841651253402233, + 0.03675059974193573, + 0.009734923020005226, + 0.006962073966860771, + 0.008332304656505585, + 0.0071355910040438175, + 0.007710146717727184, + 0.007982165552675724, + 0.00810045376420021, + 0.0088814003393054, + 0.006915430072695017, + 0.0033340584486722946, + 0.006060762796550989, + 0.003593328408896923, + 0.006273679435253143, + 0.009611101821064949, + 0.008044013753533363, + 0.009043830446898937, + 0.010621147230267525, + 0.006486895959824324, + 0.007302226033061743, + 0.005113753490149975, + 0.00859416089951992, + 0.11367134749889374, + 0.020036352798342705, + 0.09459976851940155, + 0.004983330611139536, + 0.006920509040355682, + 0.007432375568896532, + 0.006620991975069046, + 0.005667237564921379, + 0.011501972563564777, + 0.00958747137337923, + 0.008832171559333801, + 0.0051943277940154076, + 0.006879229564219713, + 0.005136762745678425, + 0.007449021562933922, + 0.0044747586362063885, + 0.009904584847390652, + 0.010776607319712639, + 0.0075886547565460205, + 0.0013803925830870867, + 0.0021599449682980776, + 0.0011213915422558784, + 0.001676481799222529, + 0.0020066616125404835, + 0.0034067926462739706, + 0.0020003628451377153, + 0.00242232927121222, + 0.002308432711288333, + 0.002603889675810933, + 0.003067990066483617, + 0.0018799761310219765, + 0.0030974934343248606, + 0.003330581821501255, + 0.002600850770249963, + 0.002925693755969405, + 0.003481579478830099, + 0.0021922511514276266, + 0.003148886142298579, + 0.001965330448001623, + 0.002726066391915083, + 0.022579941898584366, + 0.002492521656677127, + 0.019637856632471085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.04278075322508812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.004141089040786028 + ], + "opposite_logprob_tokens": [ + -5.504140853881836 + ], + "logprob_reward": -0.004141089040786028, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004141089040786028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14894895255565643, + 0.1038324385881424, + 0.09669430553913116, + 0.21672165393829346, + 0.2497130185365677, + 0.07240570336580276, + 0.05846506357192993, + 0.00812877994030714, + 0.011817127466201782, + 0.03327297419309616 + ], + "credit_map": [ + 0.02478037215769291, + 0.015208698809146881, + 0.01642521470785141, + 0.013618966564536095, + 0.011327582411468029, + 0.02116851881146431, + 0.017196403816342354, + 0.029223188757896423, + 0.016163531690835953, + 0.007687809877097607, + 0.011360411532223225, + 0.010641788132488728, + 0.015528733842074871, + 0.013101676478981972, + 0.011093326844274998, + 0.01825515553355217, + 0.01515695545822382, + 0.006055839359760284, + 0.00976637750864029, + 0.007012806832790375, + 0.009415040723979473, + 0.013981431722640991, + 0.012535681948065758, + 0.022770162671804428, + 0.030599823221564293, + 0.010990244336426258, + 0.014301509596407413, + 0.010431197471916676, + 0.02877506986260414, + 0.04255811870098114, + 0.02429063990712166, + 0.05477504804730415, + 0.009169111959636211, + 0.006201662123203278, + 0.006325931288301945, + 0.004411486908793449, + 0.005918225739151239, + 0.10818982869386673, + 0.018831877037882805, + 0.09066488593816757, + 0.00946889165788889, + 0.006289053708314896, + 0.007189780939370394, + 0.006687441375106573, + 0.00599161209538579, + 0.014643537811934948, + 0.010282551869750023, + 0.011852838099002838, + 0.006882522255182266, + 0.005046859849244356, + 0.0035957274958491325, + 0.005284259095788002, + 0.0037686522118747234, + 0.01715686544775963, + 0.008869619108736515, + 0.007860559970140457, + 0.0006879328866489232, + 0.0008480394026264548, + 0.0007800546009093523, + 0.0008110878989100456, + 0.0006796085508540273, + 0.0017233244143426418, + 0.0015351810725405812, + 0.001063551171682775, + 0.0014370704302564263, + 0.0010970106814056635, + 0.0009686764678917825, + 0.0012926794588565826, + 0.0010644955327734351, + 0.0025867330841720104, + 0.001833191723562777, + 0.0015372697962448, + 0.0029356819577515125, + 0.0030820679385215044, + 0.0033844183199107647, + 0.0026662969030439854, + 0.004083133768290281, + 0.005229054484516382, + 0.007649644277989864, + 0.004242678172886372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.004141089040786028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.07006672024726868 + ], + "opposite_logprob_tokens": [ + -2.6950666904449463 + ], + "logprob_reward": -0.07006672024726868, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006672024726868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17982245981693268, + 0.05690060555934906, + 0.1434767097234726, + 0.09306393563747406, + 0.12709841132164001, + 0.18163256347179413, + 0.1119007095694542, + 0.023017454892396927, + 0.0366922952234745, + 0.04639483615756035 + ], + "credit_map": [ + 0.019352499395608902, + 0.026445508003234863, + 0.021262243390083313, + 0.017007604241371155, + 0.01926189474761486, + 0.023162342607975006, + 0.022740615531802177, + 0.030589750036597252, + 0.01204007863998413, + 0.005295847076922655, + 0.00566417071968317, + 0.004147489555180073, + 0.004505768418312073, + 0.007256299257278442, + 0.0077498964965343475, + 0.010241052135825157, + 0.022165928035974503, + 0.012926542200148106, + 0.02074773795902729, + 0.012802657671272755, + 0.014157422818243504, + 0.018762769177556038, + 0.02296299673616886, + 0.018950646743178368, + 0.016725758090615273, + 0.005000251345336437, + 0.006114332936704159, + 0.00499249342828989, + 0.009020112454891205, + 0.019873376935720444, + 0.010495299473404884, + 0.020842308178544044, + 0.0093076778575778, + 0.004864910151809454, + 0.008531128987669945, + 0.008018072694540024, + 0.008951191790401936, + 0.040879279375076294, + 0.011209801770746708, + 0.03533635288476944, + 0.019638001918792725, + 0.01762496307492256, + 0.019406650215387344, + 0.02022312767803669, + 0.015945788472890854, + 0.028692783787846565, + 0.03325532376766205, + 0.026845941320061684, + 0.010021301917731762, + 0.012274199165403843, + 0.009267483837902546, + 0.012505580671131611, + 0.008088134229183197, + 0.02352825179696083, + 0.020957615226507187, + 0.01525813713669777, + 0.0023890710435807705, + 0.001863188692368567, + 0.0017481126124039292, + 0.0021628120448440313, + 0.0016777589917182922, + 0.006141915451735258, + 0.003029207931831479, + 0.004005388356745243, + 0.004947380628436804, + 0.0034019825980067253, + 0.005151283461600542, + 0.004641338251531124, + 0.004310975782573223, + 0.004550260957330465, + 0.005397879518568516, + 0.00429119635373354, + 0.00458642328158021, + 0.0030558134894818068, + 0.003828018205240369, + 0.0033938679844141006, + 0.00456117233261466, + 0.013546949252486229, + 0.0034425260964781046, + 0.00998006947338581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.07006672024726868 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.0019469369435682893 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.0019469369435682893, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019469369435682893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11798864603042603, + 0.19513124227523804, + 0.06278659403324127, + 0.07582452893257141, + 0.1758561134338379, + 0.10123709589242935, + 0.11919820308685303, + 0.025303972885012627, + 0.024631667882204056, + 0.10204193741083145 + ], + "credit_map": [ + 0.015928814187645912, + 0.015317490324378014, + 0.01799660362303257, + 0.01158530730754137, + 0.00907120294868946, + 0.01625569723546505, + 0.012104129418730736, + 0.01972939819097519, + 0.02792922966182232, + 0.01892528124153614, + 0.028161393478512764, + 0.02155156247317791, + 0.022078953683376312, + 0.024705419316887856, + 0.02396598644554615, + 0.02781342715024948, + 0.007521442603319883, + 0.005335894878953695, + 0.0075566768646240234, + 0.007866975851356983, + 0.00793970562517643, + 0.007444382179528475, + 0.008566326461732388, + 0.010555187240242958, + 0.00929559487849474, + 0.004286773502826691, + 0.005876213312149048, + 0.0045586139895021915, + 0.01042281836271286, + 0.01489263866096735, + 0.010712241753935814, + 0.015779636800289154, + 0.007654066197574139, + 0.004838976077735424, + 0.0048581939190626144, + 0.0031104872468858957, + 0.00587472366169095, + 0.0755799412727356, + 0.012284890748560429, + 0.06165483593940735, + 0.012395072728395462, + 0.008224551565945148, + 0.009181815199553967, + 0.008623168803751469, + 0.007659029681235552, + 0.023874083533883095, + 0.013170982711017132, + 0.018108392134308815, + 0.011806677095592022, + 0.009249740280210972, + 0.00846137385815382, + 0.01220406498759985, + 0.008610627613961697, + 0.0359058678150177, + 0.016913864761590958, + 0.016045989468693733, + 0.0023688164073973894, + 0.0026042566169053316, + 0.002037718892097473, + 0.0033037709072232246, + 0.00240239012055099, + 0.00354938767850399, + 0.006429122295230627, + 0.002608510432764888, + 0.002692857291549444, + 0.002440644195303321, + 0.0034997507464140654, + 0.003313022432848811, + 0.0023403556551784277, + 0.0034784721210598946, + 0.003199385479092598, + 0.0036671808920800686, + 0.007854086346924305, + 0.007135809399187565, + 0.008097203448414803, + 0.0064822230488061905, + 0.008431690745055676, + 0.022737620398402214, + 0.022209782153367996, + 0.019093520939350128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.0019469369435682893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371192932128906 + ], + "opposite_logprob_tokens": [ + -0.38711923360824585 + ], + "logprob_reward": -1.1371192932128906, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.1371192932128906, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1806916892528534, + 0.10497567802667618, + 0.0831545814871788, + 0.13061828911304474, + 0.36733025312423706, + 0.04147559031844139, + 0.040695060044527054, + 0.02134351246058941, + 0.015352863818407059, + 0.014362410642206669 + ], + "credit_map": [ + 0.027579370886087418, + 0.02237607352435589, + 0.018930558115243912, + 0.01865229196846485, + 0.01469742227345705, + 0.028119470924139023, + 0.02169274352490902, + 0.028643768280744553, + 0.014554244466125965, + 0.010073147714138031, + 0.017933188006281853, + 0.009725913405418396, + 0.014567028731107712, + 0.0130464443936944, + 0.012324905022978783, + 0.012750823982059956, + 0.012280425056815147, + 0.007498328574001789, + 0.009361961856484413, + 0.012483696453273296, + 0.007277187891304493, + 0.009548542089760303, + 0.012422268278896809, + 0.012282178737223148, + 0.016947120428085327, + 0.007545165251940489, + 0.0084103774279356, + 0.00792971346527338, + 0.012249520048499107, + 0.03157041594386101, + 0.014295293018221855, + 0.0316707007586956, + 0.016360972076654434, + 0.010181289166212082, + 0.010991382412612438, + 0.006797743495553732, + 0.011075837537646294, + 0.15778748691082, + 0.026187997311353683, + 0.12794755399227142, + 0.0034950547851622105, + 0.00393141433596611, + 0.0036472633946686983, + 0.005055711138993502, + 0.003535784315317869, + 0.00782418716698885, + 0.007953300140798092, + 0.006032878998667002, + 0.006235530134290457, + 0.005049300380051136, + 0.0025347955524921417, + 0.003844986902549863, + 0.002637316007167101, + 0.010290192440152168, + 0.004864245653152466, + 0.005238695070147514, + 0.001860812772065401, + 0.0022206087596714497, + 0.0018983222544193268, + 0.002263664035126567, + 0.0019606701098382473, + 0.004677142016589642, + 0.0033933205995708704, + 0.0030689730774611235, + 0.001610661973245442, + 0.0015477117849513888, + 0.0017925652209669352, + 0.0019267406314611435, + 0.002009628340601921, + 0.00230590277351439, + 0.001985013484954834, + 0.002174641937017441, + 0.001049037091434002, + 0.0015086762141436338, + 0.0009142998605966568, + 0.0012975618010386825, + 0.001139446976594627, + 0.002779296599328518, + 0.0038087088614702225, + 0.0018653834704309702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -1.1371192932128906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extremely close-up, static macro shot captures a pool of smooth, dark syrup as it slowly solidifies on a cool, matte gray surface, the entire scene rendered in dramatic, high-contrast black and white. A harsh key light sweeps across the frame from a low side angle, casting deep, velvety shadows and creating brilliant, mirrored highlights across the syrup's initially viscous, mirrored surface. As the fluid edges of the liquid begin to solidify, the transformation begins, and the once perfectly smooth areas begin to fracture from within. Intricate crystalline structures emerge and branch like frost patterns, their sharp facets catching the harsh light, causing them to shimmer brilliantly white against the deep black of the solidifying substance. The camera remains perfectly still, allowing the viewer to witness the mesmerizing, almost geological process of the organic liquid's inexorable hardening, its texture transforming from smooth and glossy to a complex, shimmering, fractured solid, creating an intense, abstract atmosphere.", + "seed": 11903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_15/index_5129_seed_11903_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": -0.4149312873836607, + "vlm_logprob": -0.4149312873836607, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.4149312873836607, + "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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.03507944941520691 + ], + "opposite_logprob_tokens": [ + -3.5350794792175293 + ], + "logprob_reward": -0.03507944941520691, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03507944941520691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12287834286689758, + 0.2838870584964752, + 0.03474494814872742, + 0.17867884039878845, + 0.1540611833333969, + 0.07758895307779312, + 0.03910093382000923, + 0.03337176516652107, + 0.04069458320736885, + 0.034993380308151245 + ], + "credit_map": [ + 0.01912810653448105, + 0.01727205142378807, + 0.011974780820310116, + 0.012520071119070053, + 0.015805643051862717, + 0.014455273747444153, + 0.016481665894389153, + 0.015240752138197422, + 0.07207074761390686, + 0.022145787253975868, + 0.02881203219294548, + 0.018125522881746292, + 0.030895739793777466, + 0.028783787041902542, + 0.014764454215765, + 0.06828898191452026, + 0.005083828698843718, + 0.004857559688389301, + 0.0046235620975494385, + 0.00420633889734745, + 0.0030753405299037695, + 0.0042749932035803795, + 0.004823741503059864, + 0.003799583064392209, + 0.02186080627143383, + 0.06138496845960617, + 0.017788413912057877, + 0.013154953718185425, + 0.020327899605035782, + 0.010675961151719093, + 0.014189369976520538, + 0.01929645799100399, + 0.017198914662003517, + 0.027396297082304955, + 0.018663350492715836, + 0.009459328837692738, + 0.04299459978938103, + 0.013884411193430424, + 0.011950921267271042, + 0.012513350695371628, + 0.010922296904027462, + 0.018908698111772537, + 0.011545018292963505, + 0.0061523145996034145, + 0.007310198619961739, + 0.011904668994247913, + 0.004331759177148342, + 0.006513997912406921, + 0.004025779664516449, + 0.003375972853973508, + 0.007098738104104996, + 0.0031568529084324837, + 0.00453344639390707, + 0.005117196124047041, + 0.004808906000107527, + 0.006984042469412088, + 0.004044156521558762, + 0.003167686052620411, + 0.007448675576597452, + 0.0030138648580759764, + 0.0028052828274667263, + 0.0041644093580543995, + 0.0032685245387256145, + 0.0054591647349298, + 0.004835357889533043, + 0.006061490625143051, + 0.008362028747797012, + 0.0035455378238111734, + 0.006481432821601629, + 0.004745092708617449, + 0.0038923912215977907, + 0.0027712529990822077, + 0.0034763908479362726, + 0.0052158716134727, + 0.0036602800246328115, + 0.002347333822399378, + 0.008119439706206322, + 0.004872157704085112, + 0.004785159137099981, + 0.002516745822504163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.03507944941520691 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.20166251063346863 + ], + "opposite_logprob_tokens": [ + -1.701662540435791 + ], + "logprob_reward": -0.20166251063346863, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20166251063346863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2093971222639084, + 0.26306235790252686, + 0.16030742228031158, + 0.0820779874920845, + 0.025708787143230438, + 0.09046128392219543, + 0.02561706118285656, + 0.03713340684771538, + 0.034266307950019836, + 0.0719681829214096 + ], + "credit_map": [ + 0.03201550990343094, + 0.021422795951366425, + 0.021520275622606277, + 0.01732044108211994, + 0.019572239369153976, + 0.0334116630256176, + 0.021659215912222862, + 0.042475003749132156, + 0.07336419075727463, + 0.01607607491314411, + 0.01609693467617035, + 0.020923225209116936, + 0.01910988613963127, + 0.02491086721420288, + 0.013916172087192535, + 0.07866505533456802, + 0.021634504199028015, + 0.02464856207370758, + 0.019301801919937134, + 0.019150510430336, + 0.01790282502770424, + 0.020031489431858063, + 0.019906776025891304, + 0.017730968073010445, + 0.009628263302147388, + 0.02262013591825962, + 0.011078514158725739, + 0.005471815355122089, + 0.012546030804514885, + 0.00772553775459528, + 0.006420135032385588, + 0.006587568204849958, + 0.003190463874489069, + 0.004250523634254932, + 0.003663303330540657, + 0.0017065420979633927, + 0.005511023569852114, + 0.0025967792607843876, + 0.002433890476822853, + 0.0023562649730592966, + 0.013672823086380959, + 0.022040367126464844, + 0.015934694558382034, + 0.008596577681601048, + 0.007568304892629385, + 0.009731074795126915, + 0.005017710383981466, + 0.007899738848209381, + 0.004140751902014017, + 0.0023506213910877705, + 0.004327650647610426, + 0.002472627442330122, + 0.0028692421037703753, + 0.002663423540070653, + 0.0032992358319461346, + 0.0034935111179947853, + 0.0051427315920591354, + 0.004517868161201477, + 0.006415169686079025, + 0.006304670590907335, + 0.00396478408947587, + 0.004029061645269394, + 0.003659654874354601, + 0.0030994724947959185, + 0.00463722413405776, + 0.0051999702118337154, + 0.0077363052405416965, + 0.0025722873397171497, + 0.0040890248492360115, + 0.0034027022775262594, + 0.004232295788824558, + 0.0023965027648955584, + 0.004516378976404667, + 0.009143631905317307, + 0.004556834232062101, + 0.003894191700965166, + 0.02607010491192341, + 0.006012401543557644, + 0.0049169897101819515, + 0.012857663445174694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.20166251063346863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.02982724830508232 + ], + "opposite_logprob_tokens": [ + -3.529827356338501 + ], + "logprob_reward": -0.02982724830508232, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982724830508232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380850076675415, + 0.13109147548675537, + 0.06898654252290726, + 0.21203172206878662, + 0.05408899486064911, + 0.07960359752178192, + 0.052479080855846405, + 0.07742059975862503, + 0.060462918132543564, + 0.1257501095533371 + ], + "credit_map": [ + 0.019055169075727463, + 0.024839788675308228, + 0.014667605049908161, + 0.01135469414293766, + 0.02048274502158165, + 0.019505036994814873, + 0.012049175798892975, + 0.01613079011440277, + 0.030095025897026062, + 0.008511878550052643, + 0.012980859726667404, + 0.01003659050911665, + 0.00854023452848196, + 0.012205499224364758, + 0.010220631025731564, + 0.03850075602531433, + 0.014132057316601276, + 0.006477221380919218, + 0.008112871088087559, + 0.005019199103116989, + 0.005228450987488031, + 0.0074213710613548756, + 0.007328602019697428, + 0.015266766771674156, + 0.021582908928394318, + 0.06947533786296844, + 0.024729356169700623, + 0.02058078721165657, + 0.02411779947578907, + 0.010779032483696938, + 0.018527619540691376, + 0.02223888412117958, + 0.008318583481013775, + 0.007207996677607298, + 0.006951803807169199, + 0.0031114397570490837, + 0.013696488924324512, + 0.005345543380826712, + 0.0038050771690905094, + 0.0056520625948905945, + 0.009996944107115269, + 0.023789798840880394, + 0.012639087624847889, + 0.006810304708778858, + 0.006515633314847946, + 0.00905254390090704, + 0.0042022778652608395, + 0.006597007159143686, + 0.006521155126392841, + 0.004917534999549389, + 0.01074891909956932, + 0.005349299404770136, + 0.006089035887271166, + 0.006103150080889463, + 0.006219135597348213, + 0.006530848331749439, + 0.022621024399995804, + 0.006767412647604942, + 0.014599703252315521, + 0.004555645864456892, + 0.004094434902071953, + 0.0050268047489225864, + 0.007538109086453915, + 0.012217460200190544, + 0.006726589519530535, + 0.010572856292128563, + 0.012069018557667732, + 0.006316033191978931, + 0.0068214815109968185, + 0.007421739399433136, + 0.006500703282654285, + 0.0040344903245568275, + 0.007698908448219299, + 0.01658758707344532, + 0.007424339652061462, + 0.006566212512552738, + 0.04674401879310608, + 0.010474476031959057, + 0.009242117404937744, + 0.02101246267557144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.02982724830508232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.0032005298417061567 + ], + "opposite_logprob_tokens": [ + -5.753200531005859 + ], + "logprob_reward": -0.0032005298417061567, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032005298417061567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05110805109143257, + 0.1060551181435585, + 0.0359322726726532, + 0.05812535062432289, + 0.059717901051044464, + 0.04878632351756096, + 0.0513356551527977, + 0.07337512820959091, + 0.19938939809799194, + 0.3161747455596924 + ], + "credit_map": [ + 0.00830228440463543, + 0.007242529187351465, + 0.004825208336114883, + 0.004159767180681229, + 0.005447095260024071, + 0.008131850510835648, + 0.005082985386252403, + 0.007916334085166454, + 0.021102778613567352, + 0.008118501864373684, + 0.009615388698875904, + 0.009137528017163277, + 0.007259270641952753, + 0.013360723853111267, + 0.007838106714189053, + 0.029622819274663925, + 0.009409523569047451, + 0.004238992929458618, + 0.004546965938061476, + 0.0030267098918557167, + 0.0023103333078324795, + 0.0034645230043679476, + 0.002857666462659836, + 0.0060775610618293285, + 0.009501850232481956, + 0.0077132899314165115, + 0.011724820360541344, + 0.0061823520809412, + 0.006469950545579195, + 0.0062657371163368225, + 0.005222279578447342, + 0.005045067984610796, + 0.005730939097702503, + 0.011813068762421608, + 0.00708851357921958, + 0.003739922773092985, + 0.015574107877910137, + 0.005097934976220131, + 0.00557875307276845, + 0.0050946613773703575, + 0.006893625482916832, + 0.01017183717340231, + 0.009946615435183048, + 0.005363392177969217, + 0.004437267780303955, + 0.004200943745672703, + 0.004192052409052849, + 0.003580587450414896, + 0.007908105850219727, + 0.003525173058733344, + 0.007715231738984585, + 0.004332809709012508, + 0.004585694521665573, + 0.00752647640183568, + 0.006326019763946533, + 0.009416143409907818, + 0.010515413247048855, + 0.007583278696984053, + 0.012846389785408974, + 0.008067947812378407, + 0.008597915060818195, + 0.008183869533240795, + 0.008251354098320007, + 0.009328958578407764, + 0.020492257550358772, + 0.029356015846133232, + 0.04330364242196083, + 0.018840812146663666, + 0.030597412958741188, + 0.025663727894425392, + 0.01794612966477871, + 0.013189414516091347, + 0.020178182050585747, + 0.04467072710394859, + 0.01905188150703907, + 0.016113312914967537, + 0.1184384897351265, + 0.027390243485569954, + 0.022798312827944756, + 0.04753359779715538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.0032005298417061567 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048866987228394 + ], + "opposite_logprob_tokens": [ + -0.17988666892051697 + ], + "logprob_reward": -1.8048866987228394, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048866987228394, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07325057685375214, + 0.12038528174161911, + 0.08215058594942093, + 0.2866067886352539, + 0.050183702260255814, + 0.04173412546515465, + 0.04733406379818916, + 0.12434510886669159, + 0.03889259323477745, + 0.13511724770069122 + ], + "credit_map": [ + 0.012922235764563084, + 0.010422181338071823, + 0.006374411750584841, + 0.005407007411122322, + 0.010575809516012669, + 0.009847680106759071, + 0.006618922110646963, + 0.011082330718636513, + 0.021174581721425056, + 0.01282582152634859, + 0.012220789678394794, + 0.009831718169152737, + 0.01643187366425991, + 0.009665118530392647, + 0.009215776808559895, + 0.029019590467214584, + 0.01893492229282856, + 0.006958333309739828, + 0.009825066663324833, + 0.007838242687284946, + 0.004456442780792713, + 0.006328342016786337, + 0.007695345673710108, + 0.020113889127969742, + 0.03010036237537861, + 0.09466543048620224, + 0.03165402263402939, + 0.023967906832695007, + 0.03249490261077881, + 0.01605311408638954, + 0.02145862951874733, + 0.03621240705251694, + 0.005784741137176752, + 0.007865061983466148, + 0.006749010644853115, + 0.002969247056171298, + 0.01564369536936283, + 0.003409912344068289, + 0.004127060063183308, + 0.0036349697038531303, + 0.006243336014449596, + 0.007587099447846413, + 0.005897834897041321, + 0.0031570931896567345, + 0.0044601354748010635, + 0.0061066607013344765, + 0.004434294067323208, + 0.003847669344395399, + 0.009577288292348385, + 0.003987688571214676, + 0.0048446813598275185, + 0.002982453443109989, + 0.0038853981532156467, + 0.005255231633782387, + 0.00523231690749526, + 0.011569006368517876, + 0.035544831305742264, + 0.01030924916267395, + 0.025026272982358932, + 0.012441466562449932, + 0.009462746791541576, + 0.00786598026752472, + 0.009798265993595123, + 0.013896297663450241, + 0.004778418689966202, + 0.006056630052626133, + 0.008004877716302872, + 0.002892296528443694, + 0.005136667285114527, + 0.00552327698096633, + 0.0035475571639835835, + 0.002952869050204754, + 0.009916059672832489, + 0.012670106254518032, + 0.010531958192586899, + 0.008309060707688332, + 0.04132756590843201, + 0.013220718130469322, + 0.011299003846943378, + 0.027842780575156212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -1.8048866987228394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extremely close-up, static macro shot captures a pool of smooth, dark syrup as it slowly solidifies on a cool, matte gray surface, the entire scene rendered in dramatic, high-contrast black and white. A harsh key light sweeps across the frame from a low side angle, casting deep, velvety shadows and creating brilliant, mirrored highlights across the syrup's initially viscous, mirrored surface. As the fluid edges of the liquid begin to solidify, the transformation begins, and the once perfectly smooth areas begin to fracture from within. Intricate crystalline structures emerge and branch like frost patterns, their sharp facets catching the harsh light, causing them to shimmer brilliantly white against the deep black of the solidifying substance. The camera remains perfectly still, allowing the viewer to witness the mesmerizing, almost geological process of the organic liquid's inexorable hardening, its texture transforming from smooth and glossy to a complex, shimmering, fractured solid, creating an intense, abstract atmosphere.", + "seed": 11904, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_15/index_5129_seed_11904_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": -0.22538019826170058, + "vlm_logprob": -0.22538019826170058, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.22538019826170058, + "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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.1738033890724182 + ], + "opposite_logprob_tokens": [ + -1.9238033294677734 + ], + "logprob_reward": -0.1738033890724182, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1738033890724182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26667091250419617, + 0.19511693716049194, + 0.05861997231841087, + 0.025846203789114952, + 0.05149788036942482, + 0.05129639059305191, + 0.04448151960968971, + 0.1581658571958542, + 0.0679558590054512, + 0.08034844696521759 + ], + "credit_map": [ + 0.029720205813646317, + 0.017234230414032936, + 0.01846061460673809, + 0.03164447844028473, + 0.09007249772548676, + 0.03663654625415802, + 0.016229595988988876, + 0.026672741398215294, + 0.040522005409002304, + 0.024946440011262894, + 0.02118573896586895, + 0.023088058456778526, + 0.023887984454631805, + 0.022459041327238083, + 0.017587358132004738, + 0.021440302953124046, + 0.006367065943777561, + 0.008032908663153648, + 0.006784022785723209, + 0.010159583762288094, + 0.007534230127930641, + 0.008267507888376713, + 0.004430451896041632, + 0.007044199388474226, + 0.0023355272132903337, + 0.003280583070591092, + 0.0020855888724327087, + 0.0026203691959381104, + 0.005679471418261528, + 0.004352639429271221, + 0.001984979957342148, + 0.003507046028971672, + 0.004208466503769159, + 0.003969870507717133, + 0.011776960454881191, + 0.006729569751769304, + 0.004896110855042934, + 0.007678743917495012, + 0.008564205840229988, + 0.0036739525385200977, + 0.004354909062385559, + 0.0043988521210849285, + 0.006916029844433069, + 0.005407497752457857, + 0.006192145403474569, + 0.006982254795730114, + 0.009026444517076015, + 0.008018259890377522, + 0.003204595297574997, + 0.0026348072569817305, + 0.005348405335098505, + 0.00819833017885685, + 0.004969296511262655, + 0.006299710366874933, + 0.009234070777893066, + 0.0045923031866550446, + 0.013013810850679874, + 0.010356670245528221, + 0.01895088329911232, + 0.030969779938459396, + 0.0286373570561409, + 0.02314206212759018, + 0.017343565821647644, + 0.015751730650663376, + 0.006894404999911785, + 0.006275303196161985, + 0.006584281101822853, + 0.018348034471273422, + 0.00810521561652422, + 0.006921045482158661, + 0.009519628249108791, + 0.005307946354150772, + 0.011774439364671707, + 0.009616202674806118, + 0.012787815183401108, + 0.012665573507547379, + 0.009912030771374702, + 0.007385412231087685, + 0.008546867407858372, + 0.007660099305212498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.1738033890724182 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.16046151518821716 + ], + "opposite_logprob_tokens": [ + -1.9104615449905396 + ], + "logprob_reward": -0.16046151518821716, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16046151518821716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06977207958698273, + 0.14702029526233673, + 0.054641060531139374, + 0.015146012417972088, + 0.06969279795885086, + 0.10891473293304443, + 0.03947713226079941, + 0.30990907549858093, + 0.10720235854387283, + 0.07822443544864655 + ], + "credit_map": [ + 0.00962419155985117, + 0.010260072536766529, + 0.006996838841587305, + 0.009865352883934975, + 0.01362990215420723, + 0.007881305180490017, + 0.0045682513155043125, + 0.006946165580302477, + 0.027065496891736984, + 0.016085641458630562, + 0.012957364320755005, + 0.02055364102125168, + 0.01798328012228012, + 0.01642700843513012, + 0.011083006858825684, + 0.024864856153726578, + 0.00740898447111249, + 0.0070336731150746346, + 0.0067191836424171925, + 0.007199751678854227, + 0.00761249428614974, + 0.007297374773770571, + 0.003942285198718309, + 0.0074273101054131985, + 0.001531902700662613, + 0.0020833013113588095, + 0.00158026535063982, + 0.002205253578722477, + 0.001881420728750527, + 0.002453747671097517, + 0.001813226263038814, + 0.0015968965599313378, + 0.006462269928306341, + 0.007015545386821032, + 0.009616037830710411, + 0.010675392113626003, + 0.011444863863289356, + 0.009148572571575642, + 0.008755125105381012, + 0.006574991159141064, + 0.011005577631294727, + 0.00959433801472187, + 0.014282234944403172, + 0.00922197476029396, + 0.022531183436512947, + 0.015556348487734795, + 0.016995014622807503, + 0.00972805730998516, + 0.002432693727314472, + 0.0025442910846322775, + 0.004511025734245777, + 0.006602937821298838, + 0.004696665331721306, + 0.005804429762065411, + 0.007553702220320702, + 0.005331388209015131, + 0.01961253583431244, + 0.021454855799674988, + 0.03004417009651661, + 0.08696974068880081, + 0.046960342675447464, + 0.04272335022687912, + 0.03750384971499443, + 0.024640245363116264, + 0.015538080595433712, + 0.02183741144835949, + 0.007915748283267021, + 0.012679871171712875, + 0.016507456079125404, + 0.01097960490733385, + 0.013636291958391666, + 0.008107897825539112, + 0.009935540147125721, + 0.008283241651952267, + 0.007649692706763744, + 0.011784273199737072, + 0.010541814379394054, + 0.01137738861143589, + 0.011818738654255867, + 0.006833745166659355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.16046151518821716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.033731408417224884 + ], + "opposite_logprob_tokens": [ + -3.408731460571289 + ], + "logprob_reward": -0.033731408417224884, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033731408417224884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1356751173734665, + 0.10962612926959991, + 0.19041204452514648, + 0.06490068882703781, + 0.17974159121513367, + 0.09297643601894379, + 0.02953312173485756, + 0.1161293312907219, + 0.046421293169260025, + 0.03458424285054207 + ], + "credit_map": [ + 0.014481802470982075, + 0.00864653754979372, + 0.007972517050802708, + 0.017510656267404556, + 0.049201805144548416, + 0.017622871324419975, + 0.007622053846716881, + 0.01261687558144331, + 0.01665302738547325, + 0.017578065395355225, + 0.01120832934975624, + 0.014947570860385895, + 0.014417226426303387, + 0.012908277101814747, + 0.008378094993531704, + 0.01353553868830204, + 0.0326717235147953, + 0.02056039124727249, + 0.019522402435541153, + 0.046977270394563675, + 0.019929442554712296, + 0.02533283270895481, + 0.014965577982366085, + 0.0104523915797472, + 0.008998810313642025, + 0.007603838108479977, + 0.005767814815044403, + 0.011750356294214725, + 0.010599285364151001, + 0.010774284601211548, + 0.005314810201525688, + 0.0040914881974458694, + 0.01919061876833439, + 0.017549509182572365, + 0.0430634506046772, + 0.020045440644025803, + 0.016844019293785095, + 0.020476100966334343, + 0.02850559912621975, + 0.014066854491829872, + 0.005740689113736153, + 0.009970867075026035, + 0.008882560767233372, + 0.008522415533661842, + 0.023912256583571434, + 0.015023852698504925, + 0.011304237879812717, + 0.009619553573429585, + 0.0028175089973956347, + 0.0023233527317643166, + 0.0039484878070652485, + 0.0057216850109398365, + 0.004266167059540749, + 0.0037548395339399576, + 0.003122398629784584, + 0.0035786833614110947, + 0.006697121076285839, + 0.007910036481916904, + 0.012725402601063251, + 0.031937871128320694, + 0.015478218905627728, + 0.015032965689897537, + 0.016543792560696602, + 0.009803920984268188, + 0.004405537154525518, + 0.00437667453661561, + 0.004798997193574905, + 0.007447624113410711, + 0.006788650061935186, + 0.005450024269521236, + 0.007377299014478922, + 0.005776488687843084, + 0.004133922979235649, + 0.0044934386387467384, + 0.0038991908077150583, + 0.006066536996513605, + 0.0034928841050714254, + 0.004365041386336088, + 0.0044675846584141254, + 0.0036656418815255165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.033731408417224884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": [ + -0.0010518262861296535 + ], + "opposite_logprob_tokens": [ + -6.876051902770996 + ], + "logprob_reward": -0.0010518262861296535, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010518262861296535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1279795914888382, + 0.2045724242925644, + 0.06789366900920868, + 0.025884510949254036, + 0.0694744661450386, + 0.09022463113069534, + 0.1174684464931488, + 0.11358898133039474, + 0.1355026811361313, + 0.0474105030298233 + ], + "credit_map": [ + 0.014867997728288174, + 0.014525831677019596, + 0.013056764379143715, + 0.014613194391131401, + 0.03158002346754074, + 0.02013068087399006, + 0.009450234472751617, + 0.009754873812198639, + 0.0266769640147686, + 0.03313322737812996, + 0.027387740090489388, + 0.030499210581183434, + 0.028118234127759933, + 0.025233404710888863, + 0.015127912163734436, + 0.018395736813545227, + 0.009670853614807129, + 0.007582644931972027, + 0.008719343692064285, + 0.008425561711192131, + 0.01028358656913042, + 0.011418335139751434, + 0.004524949472397566, + 0.0072684031911194324, + 0.0022289452608674765, + 0.0026775384321808815, + 0.00222645397298038, + 0.0027092015370726585, + 0.006239668000489473, + 0.005954441614449024, + 0.0018316206987947226, + 0.002016640966758132, + 0.006460100878030062, + 0.008478918112814426, + 0.008278527297079563, + 0.01145460270345211, + 0.0105390390381217, + 0.007967833429574966, + 0.009081266820430756, + 0.007214185781776905, + 0.006827069446444511, + 0.008372812531888485, + 0.00840787310153246, + 0.008383067324757576, + 0.025149499997496605, + 0.014818768948316574, + 0.010304675437510014, + 0.00796087458729744, + 0.007584294769912958, + 0.006511793937534094, + 0.013851712457835674, + 0.02057041972875595, + 0.014160171151161194, + 0.016947252675890923, + 0.02651379257440567, + 0.011329015716910362, + 0.01173239666968584, + 0.009755091741681099, + 0.015507069416344166, + 0.02127368375658989, + 0.017374863848090172, + 0.015138121321797371, + 0.012519635260105133, + 0.010288131423294544, + 0.012808898463845253, + 0.01375889964401722, + 0.01493686344474554, + 0.017147857695817947, + 0.018220743164420128, + 0.01683303527534008, + 0.026846785098314285, + 0.014949606731534004, + 0.0055176508612930775, + 0.0054168738424777985, + 0.005596404895186424, + 0.005540915764868259, + 0.007563017308712006, + 0.005719189997762442, + 0.0065065105445683, + 0.005549941677600145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0010518262861296535 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 extremely close-up macro view of a pool of dark syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a harsh key light sweep across the frame from a low side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the syrup transform from a smooth, glossy liquid into a fractured solid with crystalline structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera perfectly still 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578528523445129 + ], + "opposite_logprob_tokens": [ + -0.6328528523445129 + ], + "logprob_reward": -0.7578528523445129, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578528523445129, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06832513958215714, + 0.0885186642408371, + 0.12997695803642273, + 0.049763526767492294, + 0.07956225425004959, + 0.16534407436847687, + 0.07702871412038803, + 0.10205302387475967, + 0.08155670762062073, + 0.15787096321582794 + ], + "credit_map": [ + 0.010898957028985023, + 0.005834842566400766, + 0.008176257833838463, + 0.009230440482497215, + 0.010456477291882038, + 0.008248716592788696, + 0.006721349898725748, + 0.008758092299103737, + 0.0113906841725111, + 0.014486276544630527, + 0.007771784905344248, + 0.014921719208359718, + 0.012099391780793667, + 0.010528641752898693, + 0.0067983162589371204, + 0.010521842166781425, + 0.014795424416661263, + 0.01465365756303072, + 0.015598147176206112, + 0.018631981685757637, + 0.01755845732986927, + 0.01881496049463749, + 0.011869180016219616, + 0.0180551465600729, + 0.004857735708355904, + 0.004835926927626133, + 0.006398077122867107, + 0.010249905288219452, + 0.008476067334413528, + 0.00610398268327117, + 0.005470186937600374, + 0.003371646162122488, + 0.008266218937933445, + 0.011318979784846306, + 0.007567266933619976, + 0.0106171490624547, + 0.012160367332398891, + 0.01091159787029028, + 0.01128669548779726, + 0.007433973252773285, + 0.010449814610183239, + 0.015236574225127697, + 0.020958952605724335, + 0.015741616487503052, + 0.0374952033162117, + 0.02525285817682743, + 0.026312021538615227, + 0.013897046446800232, + 0.004441395401954651, + 0.004891870077699423, + 0.009176849387586117, + 0.009034116752445698, + 0.009859454818069935, + 0.011882317252457142, + 0.01974668726325035, + 0.007996022701263428, + 0.00722160330042243, + 0.009146785363554955, + 0.008060669526457787, + 0.02327643893659115, + 0.020382357761263847, + 0.015387161634862423, + 0.010929885320365429, + 0.007648117374628782, + 0.009975600056350231, + 0.012289388105273247, + 0.00892461184412241, + 0.008404186926782131, + 0.011553297750651836, + 0.00938444398343563, + 0.013597437180578709, + 0.007427747827023268, + 0.029411714524030685, + 0.017197385430336, + 0.014828415587544441, + 0.02571285143494606, + 0.018474726006388664, + 0.022299040108919144, + 0.01903379149734974, + 0.010913039557635784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.7578528523445129 + } + ], + "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/0000110/group_2/index_tensor([579, 579])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_tensor([579, 579])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..907a023ff0598da5b6fdb79abd9b59031ca6d361 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_tensor([579, 579])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, static, low-angle shot captures a life-size sculpture of a male athlete, meticulously carved from pristine white Carrara marble with subtle gray veining, perched on a simple rectangular pedestal in a minimalist art gallery. The classically styled sculpture depicts a discus thrower at the peak of his effort, his chiseled muscles tensed with frozen energy, his face a mask of concentration. The gallery's environment is stark and cool, with high white walls and a highly polished, dark gray concrete floor that reflects the focused track lighting from the ceiling, casting long, dramatic shadows. Suddenly, with a sharp crack, the marble statue animates, launching itself forward from its pedestal in a powerful, impossible leap. In a breathtaking moment, it hovers in the air before descending swiftly and decisively, landing with a deep, resounding thud that echoes throughout the vast space. Upon impact, a fine cloud of concrete and marble dust erupts from its feet, and spiderweb-like cracks immediately radiate onto the polished floor, shattering the pristine surface and creating a tense, surreal atmosphere of raw power and shattered silence.", + "seed": 31341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_579_seed_31341_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": -0.055296760005876425, + "vlm_logprob": -0.055296760005876425, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.055296760005876425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.24127303063869476 + ], + "opposite_logprob_tokens": [ + -1.741273045539856 + ], + "logprob_reward": -0.24127303063869476, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.24127303063869476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15941454470157623, + 0.04027383774518967, + 0.058801837265491486, + 0.08370395004749298, + 0.01614070124924183, + 0.0653465986251831, + 0.08338040858507156, + 0.06302783638238907, + 0.06836564093828201, + 0.3615446388721466 + ], + "credit_map": [ + 0.01678268238902092, + 0.01629151776432991, + 0.0254072155803442, + 0.02226855978369713, + 0.01202286034822464, + 0.0202946774661541, + 0.016612574458122253, + 0.029734471812844276, + 0.0032141958363354206, + 0.004485643934458494, + 0.006989535875618458, + 0.005919727962464094, + 0.003696280997246504, + 0.005339142866432667, + 0.004125228151679039, + 0.0065040807239711285, + 0.004206372890621424, + 0.004765337333083153, + 0.017561426386237144, + 0.004096281714737415, + 0.0026457440108060837, + 0.018054958432912827, + 0.002869122428819537, + 0.00460259011015296, + 0.005578383803367615, + 0.006696981377899647, + 0.026211457327008247, + 0.006106982938945293, + 0.004063417669385672, + 0.02153138443827629, + 0.0038170989137142897, + 0.009698242880403996, + 0.002252341015264392, + 0.0013022527564316988, + 0.0022009077947586775, + 0.002911584684625268, + 0.0015944408951327205, + 0.0018300488591194153, + 0.002224264433607459, + 0.0018248602282255888, + 0.006685351021587849, + 0.0067261094227433205, + 0.00696447491645813, + 0.012288510799407959, + 0.009561237879097462, + 0.00751519575715065, + 0.007740578148514032, + 0.007865138351917267, + 0.007697954308241606, + 0.014423681423068047, + 0.008185287937521935, + 0.00845731608569622, + 0.016620662063360214, + 0.011238081380724907, + 0.008142050355672836, + 0.008615377359092236, + 0.009245119988918304, + 0.01026009488850832, + 0.007543705403804779, + 0.00766444718465209, + 0.006325544789433479, + 0.007150666322559118, + 0.006349228322505951, + 0.008489031344652176, + 0.0035624131560325623, + 0.006394695490598679, + 0.0167663786560297, + 0.006073076277971268, + 0.00639492878690362, + 0.020514704287052155, + 0.005104933865368366, + 0.003554506693035364, + 0.021229436621069908, + 0.05102461948990822, + 0.09038469195365906, + 0.025803031399846077, + 0.017743831500411034, + 0.13035337626934052, + 0.011693290434777737, + 0.013312356546521187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.24127303063869476 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.018222304061055183 + ], + "opposite_logprob_tokens": [ + -4.018222332000732 + ], + "logprob_reward": -0.018222304061055183, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018222304061055183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06798533350229263, + 0.0700082927942276, + 0.04035554081201553, + 0.02754073031246662, + 0.06248987838625908, + 0.0396416001021862, + 0.04837397113442421, + 0.0877019464969635, + 0.09428869932889938, + 0.4616139829158783 + ], + "credit_map": [ + 0.010310743935406208, + 0.007330206222832203, + 0.009504208341240883, + 0.0069916993379592896, + 0.004405578598380089, + 0.015327964909374714, + 0.006420921999961138, + 0.00769401527941227, + 0.006805085111409426, + 0.005613271147012711, + 0.01739737018942833, + 0.011838790029287338, + 0.0046033249236643314, + 0.006313685327768326, + 0.00596131756901741, + 0.011475448496639729, + 0.0038189394399523735, + 0.0031342855654656887, + 0.008359402418136597, + 0.004886448848992586, + 0.001898499089293182, + 0.008855105377733707, + 0.0028524186927825212, + 0.006550442893058062, + 0.00289096194319427, + 0.0020357549656182528, + 0.00501025328412652, + 0.00396268954500556, + 0.002878509694710374, + 0.003226185915991664, + 0.0034298328682780266, + 0.004106541164219379, + 0.008619868196547031, + 0.00477086566388607, + 0.008368045091629028, + 0.010819172486662865, + 0.005599020980298519, + 0.00833901111036539, + 0.006023855414241552, + 0.009950044564902782, + 0.003957838751375675, + 0.003882348770275712, + 0.004122728481888771, + 0.007309767417609692, + 0.0052624293603003025, + 0.0041961632668972015, + 0.005719915498048067, + 0.005190407857298851, + 0.004571836441755295, + 0.007865512743592262, + 0.0044650426134467125, + 0.005893793422728777, + 0.008739909157156944, + 0.005970093421638012, + 0.005773711483925581, + 0.005094073247164488, + 0.011539415456354618, + 0.018816044554114342, + 0.010667942464351654, + 0.008837678469717503, + 0.009817864745855331, + 0.010362605564296246, + 0.007928062230348587, + 0.009732338599860668, + 0.0072714355774223804, + 0.013825940899550915, + 0.02228528819978237, + 0.007804530672729015, + 0.007810263894498348, + 0.025646282359957695, + 0.005120366346091032, + 0.004524590447545052, + 0.02695227973163128, + 0.06577102839946747, + 0.11757194250822067, + 0.0324610099196434, + 0.02304297685623169, + 0.16920159757137299, + 0.011237168684601784, + 0.015375984832644463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.018222304061055183 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.0019564551766961813 + ], + "opposite_logprob_tokens": [ + -6.251956462860107 + ], + "logprob_reward": -0.0019564551766961813, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019564551766961813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20019955933094025, + 0.04649529233574867, + 0.04576095566153526, + 0.06334167718887329, + 0.03028540499508381, + 0.06138043850660324, + 0.12812942266464233, + 0.06344418972730637, + 0.09445088356733322, + 0.2665121555328369 + ], + "credit_map": [ + 0.019921181723475456, + 0.021337732672691345, + 0.029690204188227654, + 0.028741486370563507, + 0.015877651050686836, + 0.031110553070902824, + 0.017383769154548645, + 0.03613696247339249, + 0.0040420345030725, + 0.005065520294010639, + 0.008215948939323425, + 0.006931150332093239, + 0.00468011200428009, + 0.0053774272091686726, + 0.005521116778254509, + 0.006661982275545597, + 0.003452747128903866, + 0.00398457795381546, + 0.012507835403084755, + 0.0032831693533807993, + 0.0031911456026136875, + 0.01057426817715168, + 0.003882350167259574, + 0.004884863737970591, + 0.004725467413663864, + 0.004896730184555054, + 0.018487488850951195, + 0.005015286151319742, + 0.0040885210037231445, + 0.014253934845328331, + 0.004672565031796694, + 0.007201686967164278, + 0.0037979448679834604, + 0.0026358789764344692, + 0.0038872328586876392, + 0.004327279049903154, + 0.003458755323663354, + 0.004644100554287434, + 0.0033605077769607306, + 0.004173704888671637, + 0.0054181222803890705, + 0.006063009146600962, + 0.005251200869679451, + 0.009175178594887257, + 0.009565331973135471, + 0.006821611430495977, + 0.010616392828524113, + 0.00846959464251995, + 0.013094899244606495, + 0.02032923698425293, + 0.01558428443968296, + 0.017052609473466873, + 0.02021896466612816, + 0.017600102350115776, + 0.010173237882554531, + 0.01407608576118946, + 0.006814505439251661, + 0.006249414291232824, + 0.01154792495071888, + 0.008997454307973385, + 0.006168775726109743, + 0.009690945036709309, + 0.006852505262941122, + 0.007122670765966177, + 0.006665791384875774, + 0.00861374568194151, + 0.019790351390838623, + 0.006904046982526779, + 0.008951971307396889, + 0.02704002521932125, + 0.00817926600575447, + 0.008305691182613373, + 0.014802170917391777, + 0.037020888179540634, + 0.06408482789993286, + 0.023698246106505394, + 0.015423021279275417, + 0.09036503732204437, + 0.010754902847111225, + 0.010363055393099785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0019564551766961813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.00525260902941227 + ], + "opposite_logprob_tokens": [ + -5.255252838134766 + ], + "logprob_reward": -0.00525260902941227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525260902941227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05161204934120178, + 0.03496258333325386, + 0.020792288705706596, + 0.0805443748831749, + 0.023333149030804634, + 0.05306034907698631, + 0.09953974187374115, + 0.11928307265043259, + 0.05445028096437454, + 0.46242207288742065 + ], + "credit_map": [ + 0.006608330644667149, + 0.004654804710298777, + 0.00698827626183629, + 0.006234724074602127, + 0.006715485826134682, + 0.008370345458388329, + 0.00560468016192317, + 0.006435399875044823, + 0.003223550273105502, + 0.003134075552225113, + 0.004146160092204809, + 0.00597555423155427, + 0.0034457258880138397, + 0.00746347289532423, + 0.0030022345017641783, + 0.004571807570755482, + 0.0014922987902536988, + 0.001928925747051835, + 0.005051329731941223, + 0.001992926700040698, + 0.0014898352092131972, + 0.002918282290920615, + 0.00213689636439085, + 0.0037817955017089844, + 0.005320872645825148, + 0.005771414376795292, + 0.025292959064245224, + 0.0056166863068938255, + 0.0035464931279420853, + 0.020023975521326065, + 0.005338739138096571, + 0.009633229114115238, + 0.0025983538944274187, + 0.0017485214630141854, + 0.00261499360203743, + 0.004514526110142469, + 0.0026389688719063997, + 0.0023962268605828285, + 0.0038570056203752756, + 0.0029645527247339487, + 0.00515367416664958, + 0.0061395238153636456, + 0.006216553971171379, + 0.00931516382843256, + 0.006782442796975374, + 0.007256712764501572, + 0.005152773577719927, + 0.007043501362204552, + 0.009394603781402111, + 0.008501083590090275, + 0.015937212854623795, + 0.013844682835042477, + 0.00968522485345602, + 0.010204053483903408, + 0.013425935059785843, + 0.018546944484114647, + 0.009218164719641209, + 0.012271201238036156, + 0.026217909529805183, + 0.015991540625691414, + 0.01205537561327219, + 0.01829494535923004, + 0.01180791761726141, + 0.013426024466753006, + 0.00439408840611577, + 0.006866773124784231, + 0.009888681583106518, + 0.005137643776834011, + 0.005322589073330164, + 0.013253561221063137, + 0.005673285108059645, + 0.003913655877113342, + 0.02651996538043022, + 0.06559212505817413, + 0.11674214899539948, + 0.034856826066970825, + 0.02296043373644352, + 0.16573463380336761, + 0.013149173930287361, + 0.016866791993379593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.00525260902941227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.009779401123523712 + ], + "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": [ + 0.07879290729761124, + 0.04708843305706978, + 0.04181567206978798, + 0.11509541422128677, + 0.04815078526735306, + 0.03309738636016846, + 0.030494211241602898, + 0.08897718042135239, + 0.2735113203525543, + 0.24297663569450378 + ], + "credit_map": [ + 0.010605823248624802, + 0.007379443850368261, + 0.011957486160099506, + 0.011752530932426453, + 0.006279157940298319, + 0.009208247996866703, + 0.009248360991477966, + 0.012361862696707249, + 0.004714594688266516, + 0.004077916964888573, + 0.010335727594792843, + 0.0063772136345505714, + 0.0032981226686388254, + 0.007713173981755972, + 0.005453272722661495, + 0.005118410103023052, + 0.002900177612900734, + 0.0032450391445308924, + 0.013062944635748863, + 0.003030798165127635, + 0.002008720301091671, + 0.011880949139595032, + 0.0019439002498984337, + 0.00374314421787858, + 0.007862228900194168, + 0.007321192882955074, + 0.04226991906762123, + 0.005791928619146347, + 0.002395146992057562, + 0.030738363042473793, + 0.002429658779874444, + 0.01628696918487549, + 0.0061502777971327305, + 0.0033956796396523714, + 0.0062307692132890224, + 0.009357050992548466, + 0.003990042023360729, + 0.005647941958159208, + 0.005840715020895004, + 0.007538306526839733, + 0.0029340917244553566, + 0.004705280531197786, + 0.003055991604924202, + 0.003980268258601427, + 0.002928691916167736, + 0.005751018412411213, + 0.003434596350416541, + 0.006307447794824839, + 0.003127167234197259, + 0.003782580140978098, + 0.004078228026628494, + 0.003635847708210349, + 0.0032296294812113047, + 0.004028342664241791, + 0.0038591392803937197, + 0.0047532785683870316, + 0.011080446653068066, + 0.013913514092564583, + 0.012479620054364204, + 0.011847051791846752, + 0.008733607828617096, + 0.010492637753486633, + 0.008405901491641998, + 0.012024403549730778, + 0.011742284521460533, + 0.02878601849079132, + 0.07391838729381561, + 0.019647326320409775, + 0.0246985275298357, + 0.09743903577327728, + 0.007318824529647827, + 0.009960928000509739, + 0.014080400578677654, + 0.03629297763109207, + 0.0602325014770031, + 0.01719469577074051, + 0.01184826623648405, + 0.08781209588050842, + 0.006765406578779221, + 0.008750295266509056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.009779401123523712 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, static, low-angle shot captures a life-size sculpture of a male athlete, meticulously carved from pristine white Carrara marble with subtle gray veining, perched on a simple rectangular pedestal in a minimalist art gallery. The classically styled sculpture depicts a discus thrower at the peak of his effort, his chiseled muscles tensed with frozen energy, his face a mask of concentration. The gallery's environment is stark and cool, with high white walls and a highly polished, dark gray concrete floor that reflects the focused track lighting from the ceiling, casting long, dramatic shadows. Suddenly, with a sharp crack, the marble statue animates, launching itself forward from its pedestal in a powerful, impossible leap. In a breathtaking moment, it hovers in the air before descending swiftly and decisively, landing with a deep, resounding thud that echoes throughout the vast space. Upon impact, a fine cloud of concrete and marble dust erupts from its feet, and spiderweb-like cracks immediately radiate onto the polished floor, shattering the pristine surface and creating a tense, surreal atmosphere of raw power and shattered silence.", + "seed": 31342, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_579_seed_31342_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": -0.040501737408339976, + "vlm_logprob": -0.040501737408339976, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.040501737408339976, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.16207242012023926 + ], + "opposite_logprob_tokens": [ + -2.0370724201202393 + ], + "logprob_reward": -0.16207242012023926, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.16207242012023926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22463469207286835, + 0.24837124347686768, + 0.045806821435689926, + 0.06663071364164352, + 0.059939682483673096, + 0.046261467039585114, + 0.12348395586013794, + 0.11570603400468826, + 0.03646547347307205, + 0.032699890434741974 + ], + "credit_map": [ + 0.018501754850149155, + 0.05208475887775421, + 0.015009270049631596, + 0.022307589650154114, + 0.01931883953511715, + 0.05703753978013992, + 0.022901497781276703, + 0.017473433166742325, + 0.010360172018408775, + 0.0889691710472107, + 0.00572730228304863, + 0.011810451745986938, + 0.012970592826604843, + 0.09591183066368103, + 0.011770788580179214, + 0.0108509361743927, + 0.002547335810959339, + 0.012412982061505318, + 0.003229450434446335, + 0.0032134917564690113, + 0.003519479651004076, + 0.014374042861163616, + 0.004198324866592884, + 0.0023117142263799906, + 0.0037810029461979866, + 0.02077564038336277, + 0.0024611162953078747, + 0.0035545944701880217, + 0.006865345872938633, + 0.021412024274468422, + 0.004915353376418352, + 0.0028656322974711657, + 0.006719012279063463, + 0.006270106416195631, + 0.006152996327728033, + 0.013124249875545502, + 0.0061562033370137215, + 0.005606053397059441, + 0.011280262842774391, + 0.004630799405276775, + 0.0058623929508030415, + 0.0050247302278876305, + 0.0059689711779356, + 0.009210821241140366, + 0.004435532260686159, + 0.00541390385478735, + 0.007206184323877096, + 0.0031389310024678707, + 0.0067990440875291824, + 0.009677237831056118, + 0.011160220019519329, + 0.006666004192084074, + 0.012669251300394535, + 0.03389732539653778, + 0.009089910425245762, + 0.03352496773004532, + 0.015502570196986198, + 0.023644164204597473, + 0.012521201744675636, + 0.008719957433640957, + 0.012672815471887589, + 0.01764005981385708, + 0.00905939843505621, + 0.015945862978696823, + 0.003544414881616831, + 0.0038098981603980064, + 0.00496043311432004, + 0.005335809662938118, + 0.005748428404331207, + 0.004553598817437887, + 0.004560413304716349, + 0.003952476661652327, + 0.004307769238948822, + 0.0035195532254874706, + 0.0037005373742431402, + 0.003333797911182046, + 0.00568411685526371, + 0.004962937440723181, + 0.003199255559593439, + 0.003991921432316303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.16207242012023926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.02636837214231491 + ], + "opposite_logprob_tokens": [ + -3.6513683795928955 + ], + "logprob_reward": -0.02636837214231491, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636837214231491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07876221835613251, + 0.29195940494537354, + 0.07187607884407043, + 0.18174971640110016, + 0.08126040548086166, + 0.045662958174943924, + 0.09033183008432388, + 0.05700761824846268, + 0.026700198650360107, + 0.07468953728675842 + ], + "credit_map": [ + 0.008151275105774403, + 0.017078822478652, + 0.011145386844873428, + 0.007783608511090279, + 0.005923419259488583, + 0.013564692810177803, + 0.005748383235186338, + 0.009366629645228386, + 0.012497182004153728, + 0.10977745801210403, + 0.0063165342435240746, + 0.011670101433992386, + 0.011118132621049881, + 0.1188202053308487, + 0.008981832303106785, + 0.012777956202626228, + 0.005735607817769051, + 0.017400307580828667, + 0.007561157923191786, + 0.004624662920832634, + 0.00391791108995676, + 0.02230316400527954, + 0.0051971301436424255, + 0.005136133171617985, + 0.010066484101116657, + 0.06968642771244049, + 0.0038630126509815454, + 0.006625140085816383, + 0.009615220129489899, + 0.07152918726205826, + 0.0024189199320971966, + 0.007945333607494831, + 0.008577878586947918, + 0.0073951962403953075, + 0.013019277714192867, + 0.023575274273753166, + 0.006810075603425503, + 0.007452479097992182, + 0.007567643187940121, + 0.006862573325634003, + 0.005891325883567333, + 0.0052427370101213455, + 0.0071786209009587765, + 0.005909305065870285, + 0.00609996635466814, + 0.004073469899594784, + 0.007303381338715553, + 0.003964149858802557, + 0.007616378366947174, + 0.008835405111312866, + 0.012818941846489906, + 0.005770096089690924, + 0.009425339289009571, + 0.015066965483129025, + 0.008809949271380901, + 0.021988755092024803, + 0.009860038757324219, + 0.005481468513607979, + 0.008266567252576351, + 0.004998529329895973, + 0.006824537180364132, + 0.00985213927924633, + 0.004587399773299694, + 0.007136938627809286, + 0.003597268834710121, + 0.00293947895988822, + 0.003621608018875122, + 0.002785606775432825, + 0.00458955205976963, + 0.003793128067627549, + 0.002450942061841488, + 0.002922613872215152, + 0.010919610038399696, + 0.008006319403648376, + 0.015581534244120121, + 0.005245514679700136, + 0.008966533467173576, + 0.010405834764242172, + 0.005783770699054003, + 0.00978042185306549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.02636837214231491 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.0028287656605243683 + ], + "opposite_logprob_tokens": [ + -5.877828598022461 + ], + "logprob_reward": -0.0028287656605243683, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028287656605243683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07766490429639816, + 0.1581013798713684, + 0.018768498674035072, + 0.08523159474134445, + 0.017065685242414474, + 0.052712153643369675, + 0.039782583713531494, + 0.2612650394439697, + 0.07430300116539001, + 0.21510520577430725 + ], + "credit_map": [ + 0.007708617951720953, + 0.020645180717110634, + 0.007550637703388929, + 0.008585494011640549, + 0.0045411475002765656, + 0.016067570075392723, + 0.007153217680752277, + 0.005413036327809095, + 0.006985814310610294, + 0.061286360025405884, + 0.0025912225246429443, + 0.005047813523560762, + 0.005742005072534084, + 0.0672144815325737, + 0.0026203934103250504, + 0.0066132815554738045, + 0.001572796842083335, + 0.003114116145297885, + 0.0019827866926789284, + 0.002808529417961836, + 0.002489672973752022, + 0.0025837398134171963, + 0.0024702763184905052, + 0.001746579771861434, + 0.004953079391270876, + 0.031143726781010628, + 0.0027269383426755667, + 0.003520471043884754, + 0.005108713638037443, + 0.031356874853372574, + 0.0027034312952309847, + 0.003718356601893902, + 0.0017381710931658745, + 0.0018570116953924298, + 0.0028496652375906706, + 0.0032779090106487274, + 0.002030416624620557, + 0.0018023928860202432, + 0.0018581973854452372, + 0.0016519215423613787, + 0.005019727628678083, + 0.00540425069630146, + 0.0067429495975375175, + 0.008893878199160099, + 0.005975150503218174, + 0.005602468270808458, + 0.008483869023621082, + 0.006589857861399651, + 0.005415728781372309, + 0.00601987773552537, + 0.005170509684830904, + 0.0036351089365780354, + 0.0049417815171182156, + 0.005517109762877226, + 0.0036111061926931143, + 0.005471360869705677, + 0.031130893155932426, + 0.06440278142690659, + 0.03282393887639046, + 0.013187126256525517, + 0.02418985590338707, + 0.036219485104084015, + 0.013388645835220814, + 0.045922331511974335, + 0.01150216069072485, + 0.0077828834764659405, + 0.011763343587517738, + 0.006820020265877247, + 0.011000008322298527, + 0.0077262334525585175, + 0.008501813746988773, + 0.00920653622597456, + 0.03118426539003849, + 0.023801185190677643, + 0.044730961322784424, + 0.01399240829050541, + 0.01788276433944702, + 0.03035755082964897, + 0.017554720863699913, + 0.03560134395956993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -0.0028287656605243683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.0036134920082986355 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.0036134920082986355, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036134920082986355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17401355504989624, + 0.17304877936840057, + 0.02729986235499382, + 0.16604457795619965, + 0.03261762484908104, + 0.09978365898132324, + 0.1526416838169098, + 0.09168143570423126, + 0.0376746840775013, + 0.0451941192150116 + ], + "credit_map": [ + 0.0072423252277076244, + 0.06426675617694855, + 0.0048659988678991795, + 0.007096745539456606, + 0.009164834395051003, + 0.06604094058275223, + 0.00806762557476759, + 0.007268313784152269, + 0.008092813193798065, + 0.06699203699827194, + 0.004010264296084642, + 0.0068593863397836685, + 0.006226907018572092, + 0.07072616368532181, + 0.003416884457692504, + 0.006724325940012932, + 0.002429904881864786, + 0.006375221535563469, + 0.001912567880935967, + 0.0039833299815654755, + 0.002110720844939351, + 0.005597363691776991, + 0.0031732204370200634, + 0.0017175332177430391, + 0.008146216161549091, + 0.0624108649790287, + 0.00440956512466073, + 0.00684205861762166, + 0.008105969056487083, + 0.06529650092124939, + 0.0029803458601236343, + 0.007853039540350437, + 0.0040835123509168625, + 0.0033341809175908566, + 0.005703811999410391, + 0.005303351674228907, + 0.0036447152961045504, + 0.003049295861274004, + 0.0042458451353013515, + 0.0032529118470847607, + 0.011672373861074448, + 0.008015449158847332, + 0.020778637379407883, + 0.013212217018008232, + 0.010163028724491596, + 0.010253082029521465, + 0.011523389257490635, + 0.014165478758513927, + 0.00916300155222416, + 0.014118334278464317, + 0.017676567658782005, + 0.005515695549547672, + 0.00983430165797472, + 0.04343429580330849, + 0.007079486735165119, + 0.045819997787475586, + 0.012965881265699863, + 0.01705215312540531, + 0.010532256215810776, + 0.007703911513090134, + 0.00988380890339613, + 0.01296350546181202, + 0.007661857176572084, + 0.01291806623339653, + 0.004869051743298769, + 0.004417836666107178, + 0.0055257780477404594, + 0.0038984701968729496, + 0.005925325211137533, + 0.004876868799328804, + 0.0037427712231874466, + 0.004418585449457169, + 0.006171022541821003, + 0.005375097971409559, + 0.00732822809368372, + 0.002584007801488042, + 0.006911234464496374, + 0.0065741450525820255, + 0.004041758831590414, + 0.0062086270190775394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0036134920082986355 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15567979216575623, + 0.11921572685241699, + 0.09508047252893448, + 0.3288854658603668, + 0.014114730060100555, + 0.11538827419281006, + 0.04251429811120033, + 0.06040770187973976, + 0.02851884253323078, + 0.040194716304540634 + ], + "credit_map": [ + 0.00786292739212513, + 0.05027361586689949, + 0.01013456005603075, + 0.012952500022947788, + 0.008218350820243359, + 0.051747795194387436, + 0.007361581549048424, + 0.007128463126718998, + 0.005240955390036106, + 0.044565264135599136, + 0.00295251770876348, + 0.005028745159506798, + 0.0044070095755159855, + 0.04903337359428406, + 0.0029753106646239758, + 0.005012547131627798, + 0.004869196098297834, + 0.03435051068663597, + 0.0038723009638488293, + 0.003816251875832677, + 0.004761934280395508, + 0.03573267161846161, + 0.003359902184456587, + 0.004317702259868383, + 0.015254649333655834, + 0.12556272745132446, + 0.007678532972931862, + 0.009539853781461716, + 0.015849778428673744, + 0.13496313989162445, + 0.003920004703104496, + 0.016116779297590256, + 0.0016394790727645159, + 0.0013757898705080152, + 0.0029226462356746197, + 0.003057715483009815, + 0.0011899524834007025, + 0.0012570065446197987, + 0.0012766103027388453, + 0.0013955308822914958, + 0.013254230841994286, + 0.012113488279283047, + 0.019837018102407455, + 0.012575166299939156, + 0.010160342790186405, + 0.012938260100781918, + 0.02159040793776512, + 0.012919360771775246, + 0.005202415864914656, + 0.004116948693990707, + 0.00821079034358263, + 0.0030163023620843887, + 0.00441019656136632, + 0.004638345912098885, + 0.0033117157872766256, + 0.009607584215700626, + 0.00702506210654974, + 0.013939705677330494, + 0.007130622398108244, + 0.0031700637191534042, + 0.006367016118019819, + 0.009447616524994373, + 0.0037770061753690243, + 0.009550608694553375, + 0.0038909700233489275, + 0.0033681681379675865, + 0.003841438563540578, + 0.002532160608097911, + 0.005508601199835539, + 0.0036199286114424467, + 0.002837361302226782, + 0.0029202138539403677, + 0.005501825362443924, + 0.005763135850429535, + 0.006204527337104082, + 0.0018417970277369022, + 0.007366476580500603, + 0.006211847998201847, + 0.0026802299544215202, + 0.0046248785220086575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.007625637110322714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, static, low-angle shot captures a life-size sculpture of a male athlete, meticulously carved from pristine white Carrara marble with subtle gray veining, perched on a simple rectangular pedestal in a minimalist art gallery. The classically styled sculpture depicts a discus thrower at the peak of his effort, his chiseled muscles tensed with frozen energy, his face a mask of concentration. The gallery's environment is stark and cool, with high white walls and a highly polished, dark gray concrete floor that reflects the focused track lighting from the ceiling, casting long, dramatic shadows. Suddenly, with a sharp crack, the marble statue animates, launching itself forward from its pedestal in a powerful, impossible leap. In a breathtaking moment, it hovers in the air before descending swiftly and decisively, landing with a deep, resounding thud that echoes throughout the vast space. Upon impact, a fine cloud of concrete and marble dust erupts from its feet, and spiderweb-like cracks immediately radiate onto the polished floor, shattering the pristine surface and creating a tense, surreal atmosphere of raw power and shattered silence.", + "seed": 31343, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_579_seed_31343_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": -0.07841982846148313, + "vlm_logprob": -0.07841982846148313, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.07841982846148313, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.35004061460494995 + ], + "opposite_logprob_tokens": [ + -1.3500406742095947 + ], + "logprob_reward": -0.35004061460494995, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.35004061460494995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1436414122581482, + 0.0585126131772995, + 0.05520298331975937, + 0.03261846676468849, + 0.03246886283159256, + 0.15502364933490753, + 0.18223461508750916, + 0.19435034692287445, + 0.08680853247642517, + 0.05913844704627991 + ], + "credit_map": [ + 0.016647346317768097, + 0.017893197014927864, + 0.02559276670217514, + 0.02199973724782467, + 0.015712831169366837, + 0.013418484479188919, + 0.02029169350862503, + 0.012085365131497383, + 0.007153805810958147, + 0.006213248707354069, + 0.007966912351548672, + 0.010448040440678596, + 0.006445339415222406, + 0.006156341638416052, + 0.009020720608532429, + 0.005108212120831013, + 0.008623196743428707, + 0.006963904947042465, + 0.010793701745569706, + 0.012871169485151768, + 0.0041952249594032764, + 0.004744345787912607, + 0.0037478702142834663, + 0.003263573395088315, + 0.0036971536464989185, + 0.003624711185693741, + 0.0055535053834319115, + 0.005230816546827555, + 0.0026328370440751314, + 0.0035415806341916323, + 0.0053216624073684216, + 0.003016202710568905, + 0.00423018541187048, + 0.002954888390377164, + 0.005865735001862049, + 0.005851235240697861, + 0.0022640828974545, + 0.0031632317695766687, + 0.005558009725064039, + 0.0025814957916736603, + 0.015170009806752205, + 0.01900257170200348, + 0.02293386124074459, + 0.02609916403889656, + 0.01825380139052868, + 0.016920167952775955, + 0.026136135682463646, + 0.010507950559258461, + 0.03648119792342186, + 0.024219859391450882, + 0.03631835803389549, + 0.01989753171801567, + 0.017105065286159515, + 0.01142454519867897, + 0.020328164100646973, + 0.016459926962852478, + 0.03328276053071022, + 0.01823725365102291, + 0.05744149908423424, + 0.02184363082051277, + 0.009799186140298843, + 0.013420823030173779, + 0.01438632607460022, + 0.025938883423805237, + 0.010452352464199066, + 0.00888714287430048, + 0.01283116452395916, + 0.01428734790533781, + 0.006838351953774691, + 0.006029721815139055, + 0.014155561104416847, + 0.01332689356058836, + 0.008929702453315258, + 0.005922521464526653, + 0.011576559394598007, + 0.008066664449870586, + 0.004086603876203299, + 0.0037538125179708004, + 0.00741287786513567, + 0.009389710612595081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.35004061460494995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.02061675675213337 + ], + "opposite_logprob_tokens": [ + -3.8956167697906494 + ], + "logprob_reward": -0.02061675675213337, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061675675213337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14857283234596252, + 0.054027289152145386, + 0.03021935746073723, + 0.0784965455532074, + 0.055600620806217194, + 0.07377549260854721, + 0.09243718534708023, + 0.23426026105880737, + 0.19465984404087067, + 0.03795059025287628 + ], + "credit_map": [ + 0.01413881778717041, + 0.015435205772519112, + 0.02267981320619583, + 0.026604322716593742, + 0.012747575528919697, + 0.013975049369037151, + 0.02943449281156063, + 0.013557550497353077, + 0.006824538577347994, + 0.006594664882868528, + 0.007647614926099777, + 0.010189345106482506, + 0.0045328508131206036, + 0.004332160111516714, + 0.009657183662056923, + 0.0042489259503781796, + 0.005444770213216543, + 0.0031971060670912266, + 0.004396879579871893, + 0.0051610940136015415, + 0.0026432855520397425, + 0.00259918044321239, + 0.004564197733998299, + 0.0022128436248749495, + 0.01121869869530201, + 0.008093315176665783, + 0.013269804418087006, + 0.01968267187476158, + 0.007635989226400852, + 0.004720734432339668, + 0.009867142885923386, + 0.004008186049759388, + 0.005365988239645958, + 0.005561937112361193, + 0.008050167001783848, + 0.009105167351663113, + 0.005872247274965048, + 0.0062568834982812405, + 0.010145123116672039, + 0.005243103485554457, + 0.005868889857083559, + 0.011487402021884918, + 0.008427891880273819, + 0.008776671253144741, + 0.009838638827204704, + 0.010223133489489555, + 0.012337171472609043, + 0.006815689150243998, + 0.01147396955639124, + 0.010617191903293133, + 0.012324185110628605, + 0.01230547484010458, + 0.0128055764362216, + 0.008958280086517334, + 0.011541980318725109, + 0.012410532683134079, + 0.042966749519109726, + 0.023079972714185715, + 0.07077834755182266, + 0.023513775318861008, + 0.01091755460947752, + 0.016482798382639885, + 0.016169626265764236, + 0.030351415276527405, + 0.03003118932247162, + 0.017657233402132988, + 0.06167471408843994, + 0.028620271012187004, + 0.009496665559709072, + 0.010670054703950882, + 0.019215093925595284, + 0.01729462295770645, + 0.004855705890804529, + 0.004487365484237671, + 0.005650951061397791, + 0.0050350758247077465, + 0.004267245065420866, + 0.002689648885279894, + 0.005091425497084856, + 0.005873175337910652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.02061675675213337 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.003631902625784278 + ], + "opposite_logprob_tokens": [ + -5.628632068634033 + ], + "logprob_reward": -0.003631902625784278, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003631902625784278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08919528871774673, + 0.027109825983643532, + 0.022540759295225143, + 0.0329684317111969, + 0.08119551092386246, + 0.07666146010160446, + 0.11557227373123169, + 0.18077939748764038, + 0.29194051027297974, + 0.08203662931919098 + ], + "credit_map": [ + 0.015994945541024208, + 0.008751807734370232, + 0.016264690086245537, + 0.02304225042462349, + 0.005934552755206823, + 0.004602895583957434, + 0.008133393712341785, + 0.006470745895057917, + 0.003172418801113963, + 0.003502905834466219, + 0.003682300914078951, + 0.00439488748088479, + 0.003419274929910898, + 0.002356932032853365, + 0.003992808982729912, + 0.002588295843452215, + 0.002801360795274377, + 0.002859347267076373, + 0.0031317342072725296, + 0.003744344227015972, + 0.003054784843698144, + 0.0019525923999026418, + 0.0036681536585092545, + 0.001328439568169415, + 0.005730836186558008, + 0.0032908145803958178, + 0.006020673085004091, + 0.007586812600493431, + 0.003579072654247284, + 0.0018859351985156536, + 0.003279732074588537, + 0.001594554167240858, + 0.010701654478907585, + 0.007997923530638218, + 0.013229028321802616, + 0.016439445316791534, + 0.007875881157815456, + 0.011449631303548813, + 0.006118106655776501, + 0.0073838368989527225, + 0.007336930371820927, + 0.013875680975615978, + 0.009252404794096947, + 0.010952373966574669, + 0.007743104826658964, + 0.009921624325215816, + 0.01066602859646082, + 0.006913307122886181, + 0.016794845461845398, + 0.013598640449345112, + 0.017593640834093094, + 0.012800011783838272, + 0.010073211044073105, + 0.014148268848657608, + 0.010790548287332058, + 0.019773095846176147, + 0.02856195718050003, + 0.019404947757720947, + 0.03524741157889366, + 0.024225078523159027, + 0.014312183484435081, + 0.01389433816075325, + 0.01766218990087509, + 0.02747129276394844, + 0.03326316550374031, + 0.03372061252593994, + 0.09676316380500793, + 0.053043484687805176, + 0.01802908442914486, + 0.019623976200819016, + 0.020426541566848755, + 0.017070429399609566, + 0.010097861289978027, + 0.008222157135605812, + 0.012088444083929062, + 0.014316442422568798, + 0.008443739265203476, + 0.00838850624859333, + 0.009417400695383549, + 0.011062076315283775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.003631902625784278 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.003615273628383875 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615273628383875, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615273628383875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05862518399953842, + 0.027468007057905197, + 0.01596187613904476, + 0.016558414325118065, + 0.024526843801140785, + 0.038930974900722504, + 0.04662322625517845, + 0.24782231450080872, + 0.2601313591003418, + 0.26335179805755615 + ], + "credit_map": [ + 0.008394645527005196, + 0.005739395972341299, + 0.009343056939542294, + 0.011888223700225353, + 0.00563803082332015, + 0.003982435446232557, + 0.009638877585530281, + 0.0040005166083574295, + 0.0041525173000991344, + 0.002908192342147231, + 0.004353079944849014, + 0.005368952639400959, + 0.002623345935717225, + 0.0018212017603218555, + 0.00480455718934536, + 0.0014361609937623143, + 0.0022443081252276897, + 0.0016237682430073619, + 0.002457020804286003, + 0.0034578449558466673, + 0.0018503536703065038, + 0.001295531284995377, + 0.002174471039324999, + 0.0008585777250118554, + 0.001951636397279799, + 0.0017690096283331513, + 0.0025707983877509832, + 0.0028423843905329704, + 0.0019032132113352418, + 0.0014491538750007749, + 0.003040270647034049, + 0.0010319483699277043, + 0.002697720192372799, + 0.0027191839180886745, + 0.004058598540723324, + 0.004853794816881418, + 0.0020343917421996593, + 0.002040010178461671, + 0.0046420530416071415, + 0.0014810911379754543, + 0.004402411635965109, + 0.005418365355581045, + 0.006288823205977678, + 0.007358204107731581, + 0.003534362418577075, + 0.0038488092832267284, + 0.004656639881432056, + 0.00342336087487638, + 0.006370978429913521, + 0.004780846647918224, + 0.007687246892601252, + 0.006718627642840147, + 0.0034745640587061644, + 0.004955849144607782, + 0.004803274292498827, + 0.007831837981939316, + 0.042005643248558044, + 0.0226746778935194, + 0.08544475585222244, + 0.02014731802046299, + 0.013600400649011135, + 0.015608549118041992, + 0.013512187637388706, + 0.0348287969827652, + 0.02890116721391678, + 0.03266771510243416, + 0.058521345257759094, + 0.04489721730351448, + 0.0197246503084898, + 0.01570507511496544, + 0.033997055143117905, + 0.025717148557305336, + 0.04580746218562126, + 0.020752906799316406, + 0.05639402195811272, + 0.03522925823926926, + 0.01563302055001259, + 0.01513164397329092, + 0.03874596580862999, + 0.03565750643610954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.003615273628383875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.014194594696164131 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194594696164131, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194594696164131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056339506059885025, + 0.025866633281111717, + 0.014104367233812809, + 0.017148885875940323, + 0.02204579859972, + 0.08101889491081238, + 0.2004978507757187, + 0.1648416668176651, + 0.3846414089202881, + 0.0334949791431427 + ], + "credit_map": [ + 0.0061299754306674, + 0.006511338986456394, + 0.008019616827368736, + 0.010971794836223125, + 0.00585536053404212, + 0.0050179180689156055, + 0.008657279424369335, + 0.005176222417503595, + 0.00411234050989151, + 0.0021428398322314024, + 0.004101170692592859, + 0.0053108325228095055, + 0.0023834442254155874, + 0.002635801210999489, + 0.0033404529094696045, + 0.0018397517269477248, + 0.002237979555502534, + 0.0013393429107964039, + 0.001844150247052312, + 0.002287635812535882, + 0.001441496075130999, + 0.0015798112144693732, + 0.0022642705589532852, + 0.001109680044464767, + 0.0031466037034988403, + 0.0011707632802426815, + 0.002695527859032154, + 0.0032035852782428265, + 0.0014861754607409239, + 0.0016475605079904199, + 0.0024771434254944324, + 0.001321526593528688, + 0.0032225470058619976, + 0.002169103594496846, + 0.003533212933689356, + 0.004232354462146759, + 0.0018928470090031624, + 0.002101155463606119, + 0.0032190612982958555, + 0.0016755169490352273, + 0.011100606992840767, + 0.009443294256925583, + 0.010112727992236614, + 0.01111544668674469, + 0.0065139345824718475, + 0.013725358061492443, + 0.010717070661485195, + 0.008290455676615238, + 0.02264932170510292, + 0.014509184285998344, + 0.033643439412117004, + 0.028364073485136032, + 0.01074923388659954, + 0.031362444162368774, + 0.014887208119034767, + 0.044332943856716156, + 0.027397578582167625, + 0.016955440863966942, + 0.05067391321063042, + 0.017597366124391556, + 0.009267630986869335, + 0.009897523559629917, + 0.011594622395932674, + 0.021457595750689507, + 0.057075634598731995, + 0.0308513306081295, + 0.14858178794384003, + 0.06936518102884293, + 0.014244716614484787, + 0.018945179879665375, + 0.0254046693444252, + 0.020172925665974617, + 0.0038043372333049774, + 0.0035659463610500097, + 0.005391375627368689, + 0.0049644699320197105, + 0.003136527258902788, + 0.0023942505940794945, + 0.005246467422693968, + 0.004991605877876282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.014194594696164131 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, static, low-angle shot captures a life-size sculpture of a male athlete, meticulously carved from pristine white Carrara marble with subtle gray veining, perched on a simple rectangular pedestal in a minimalist art gallery. The classically styled sculpture depicts a discus thrower at the peak of his effort, his chiseled muscles tensed with frozen energy, his face a mask of concentration. The gallery's environment is stark and cool, with high white walls and a highly polished, dark gray concrete floor that reflects the focused track lighting from the ceiling, casting long, dramatic shadows. Suddenly, with a sharp crack, the marble statue animates, launching itself forward from its pedestal in a powerful, impossible leap. In a breathtaking moment, it hovers in the air before descending swiftly and decisively, landing with a deep, resounding thud that echoes throughout the vast space. Upon impact, a fine cloud of concrete and marble dust erupts from its feet, and spiderweb-like cracks immediately radiate onto the polished floor, shattering the pristine surface and creating a tense, surreal atmosphere of raw power and shattered silence.", + "seed": 31344, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_2/index_579_seed_31344_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": -0.0399330664658919, + "vlm_logprob": -0.0399330664658919, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.0399330664658919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.13717100024223328 + ], + "opposite_logprob_tokens": [ + -2.3871710300445557 + ], + "logprob_reward": -0.13717100024223328, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.13717100024223328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18253891170024872, + 0.06777811795473099, + 0.05606859177350998, + 0.06761226058006287, + 0.0554162859916687, + 0.1248769760131836, + 0.11302240937948227, + 0.08798036724328995, + 0.2196691930294037, + 0.025036871433258057 + ], + "credit_map": [ + 0.014082476496696472, + 0.018040385097265244, + 0.013807795941829681, + 0.0296940878033638, + 0.021101748570799828, + 0.03312211111187935, + 0.02403445914387703, + 0.02865585684776306, + 0.004009841475635767, + 0.00660585705190897, + 0.0040017105638980865, + 0.01100134663283825, + 0.009178461506962776, + 0.012495415285229683, + 0.007825249806046486, + 0.012660236097872257, + 0.0028498531319200993, + 0.004983898252248764, + 0.002532048150897026, + 0.007779995910823345, + 0.009883841499686241, + 0.008009303361177444, + 0.010164642706513405, + 0.009865006431937218, + 0.002397477626800537, + 0.006437777075916529, + 0.0026455584447830915, + 0.006931427866220474, + 0.011597640812397003, + 0.009906764142215252, + 0.020621901378035545, + 0.007073717191815376, + 0.003147990209981799, + 0.005836889613419771, + 0.002538505243137479, + 0.009072668850421906, + 0.008251756429672241, + 0.008653014898300171, + 0.008747267536818981, + 0.009168190881609917, + 0.00516675366088748, + 0.01766827516257763, + 0.004661005921661854, + 0.01354474201798439, + 0.020819533616304398, + 0.0177732165902853, + 0.03246752917766571, + 0.012775917537510395, + 0.006892600562423468, + 0.00822775810956955, + 0.006499812938272953, + 0.013712339103221893, + 0.014440431259572506, + 0.01653279736638069, + 0.035973623394966125, + 0.010743042454123497, + 0.0047632562927901745, + 0.013202148489654064, + 0.005871244706213474, + 0.01570459082722664, + 0.013455052860081196, + 0.01273155678063631, + 0.01690024696290493, + 0.005352266598492861, + 0.023461254313588142, + 0.009825626388192177, + 0.006674154661595821, + 0.007158889435231686, + 0.029383620247244835, + 0.07006523758172989, + 0.00908389687538147, + 0.06401652097702026, + 0.0014752992428839207, + 0.002322156447917223, + 0.002644008956849575, + 0.0039307852275669575, + 0.003567525651305914, + 0.003165281843394041, + 0.004231494851410389, + 0.003700320143252611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.13717100024223328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.05500213801860809 + ], + "opposite_logprob_tokens": [ + -2.930002212524414 + ], + "logprob_reward": -0.05500213801860809, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500213801860809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04133952409029007, + 0.04712741822004318, + 0.030171383172273636, + 0.07517692446708679, + 0.026058046147227287, + 0.1824754923582077, + 0.06966277211904526, + 0.06892026960849762, + 0.4154799282550812, + 0.043588217347860336 + ], + "credit_map": [ + 0.002666591200977564, + 0.00418774364516139, + 0.0022918316535651684, + 0.008044753223657608, + 0.004937772173434496, + 0.007153838407248259, + 0.006471064407378435, + 0.00558593450114131, + 0.0028079389594495296, + 0.0056664603762328625, + 0.002211622428148985, + 0.0052045537158846855, + 0.008978823199868202, + 0.0073036011308431625, + 0.009423154406249523, + 0.005531268659979105, + 0.0014535440132021904, + 0.003358375746756792, + 0.0016961153596639633, + 0.003211807459592819, + 0.006174453999847174, + 0.004401903133839369, + 0.00707363523542881, + 0.0028015505522489548, + 0.0032073857728391886, + 0.006351284217089415, + 0.0036361422389745712, + 0.008495353162288666, + 0.013083447702229023, + 0.009023367427289486, + 0.019863083958625793, + 0.011516871862113476, + 0.00125968549400568, + 0.0017061425605788827, + 0.0011506379814818501, + 0.0032547516748309135, + 0.002861747983843088, + 0.0035366553347557783, + 0.008697353303432465, + 0.003591073676943779, + 0.008457236923277378, + 0.02720773033797741, + 0.006996150128543377, + 0.021122664213180542, + 0.026234624907374382, + 0.02744450233876705, + 0.04674777016043663, + 0.018264830112457275, + 0.005741126369684935, + 0.006146478466689587, + 0.004841997288167477, + 0.01595172844827175, + 0.007652363274246454, + 0.01101937610656023, + 0.009910941123962402, + 0.008398767560720444, + 0.007645810022950172, + 0.008072004653513432, + 0.0066904411651194096, + 0.009457508102059364, + 0.010576379485428333, + 0.009887287393212318, + 0.00870859157294035, + 0.007882255129516125, + 0.04440109059214592, + 0.011677103117108345, + 0.011150588281452656, + 0.01159211527556181, + 0.056400883942842484, + 0.13313354551792145, + 0.02464149333536625, + 0.12248315662145615, + 0.002581933978945017, + 0.005470103118568659, + 0.004774782340973616, + 0.008223711512982845, + 0.0045991698279976845, + 0.006188295315951109, + 0.005513925105333328, + 0.0062362998723983765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.05500213801860809 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.0010590903693810105 + ], + "opposite_logprob_tokens": [ + -6.876059055328369 + ], + "logprob_reward": -0.0010590903693810105, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010590903693810105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11105827242136002, + 0.055561937391757965, + 0.05312446504831314, + 0.08889371901750565, + 0.05309047922492027, + 0.06286446750164032, + 0.1105121374130249, + 0.25161978602409363, + 0.17593267560005188, + 0.03734205663204193 + ], + "credit_map": [ + 0.008190568536520004, + 0.01386652048677206, + 0.0077752466313540936, + 0.019335495308041573, + 0.0167812742292881, + 0.01681423932313919, + 0.011669980362057686, + 0.016624944284558296, + 0.002152610570192337, + 0.005397467873990536, + 0.0021718249190598726, + 0.006198465824127197, + 0.009711886756122112, + 0.008850384503602982, + 0.015571522526443005, + 0.0055077760480344296, + 0.0022930409759283066, + 0.004909189883619547, + 0.002219719812273979, + 0.0061612557619810104, + 0.008461407385766506, + 0.008303457871079445, + 0.015114800073206425, + 0.005661595147103071, + 0.0034717002417892218, + 0.00860885065048933, + 0.003298127790912986, + 0.010862890630960464, + 0.018654542043805122, + 0.010823225602507591, + 0.02069222368299961, + 0.012482154183089733, + 0.0025366125628352165, + 0.0070719229988753796, + 0.002151231514289975, + 0.008933025412261486, + 0.009798224084079266, + 0.007957804016768932, + 0.00816318392753601, + 0.0064784749411046505, + 0.0030328931752592325, + 0.007078418508172035, + 0.003706220770254731, + 0.008790645748376846, + 0.008782082237303257, + 0.009548967704176903, + 0.015533475205302238, + 0.006391764618456364, + 0.008234353736042976, + 0.012125286273658276, + 0.006834626663476229, + 0.017428774386644363, + 0.013735632412135601, + 0.016731474548578262, + 0.02458352781832218, + 0.010838465765118599, + 0.013627071864902973, + 0.034981898963451385, + 0.014093582518398762, + 0.044196292757987976, + 0.04195871949195862, + 0.0373048335313797, + 0.04892491176724434, + 0.016532501205801964, + 0.017283760011196136, + 0.012420807033777237, + 0.010897061787545681, + 0.010118731297552586, + 0.02195231430232525, + 0.04599841684103012, + 0.013519629836082458, + 0.04374196007847786, + 0.0028840205632150173, + 0.0040394156239926815, + 0.003171890275552869, + 0.004924207925796509, + 0.007805780507624149, + 0.005298130214214325, + 0.003901951014995575, + 0.0053166598081588745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0010590903693810105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.0028186615090817213 + ], + "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": [ + 0.04819817841053009, + 0.034120745956897736, + 0.02418949268758297, + 0.045270197093486786, + 0.04475890472531319, + 0.03674624115228653, + 0.050721678882837296, + 0.06333614140748978, + 0.5414936542510986, + 0.111164890229702 + ], + "credit_map": [ + 0.004068785812705755, + 0.005797194316983223, + 0.003599352901801467, + 0.00827642623335123, + 0.006542486138641834, + 0.00809415802359581, + 0.006300343666225672, + 0.005519422702491283, + 0.0015480213332921267, + 0.004117893520742655, + 0.0012661557411774993, + 0.003260243684053421, + 0.006083250045776367, + 0.006501193158328533, + 0.008280029520392418, + 0.003063945099711418, + 0.001129628042690456, + 0.002513175830245018, + 0.0011929366737604141, + 0.0024649964179843664, + 0.003809618763625622, + 0.004122752696275711, + 0.00640986068174243, + 0.00254651578143239, + 0.0025185232516378164, + 0.003852339694276452, + 0.002047592308372259, + 0.005683633033186197, + 0.008012712001800537, + 0.005757011938840151, + 0.009683321230113506, + 0.007715053856372833, + 0.0021706130355596542, + 0.004688417539000511, + 0.001972625497728586, + 0.006678492296487093, + 0.0058581046760082245, + 0.00607235636562109, + 0.010835196822881699, + 0.006483088713139296, + 0.0019938002806156874, + 0.007368198595941067, + 0.001397654414176941, + 0.00360772293061018, + 0.0064530204981565475, + 0.006155954208225012, + 0.004892751108855009, + 0.004877129569649696, + 0.007216311059892178, + 0.00510948384180665, + 0.00397169915959239, + 0.0053833285346627235, + 0.004931452218443155, + 0.007075545843690634, + 0.01118377223610878, + 0.005850073415786028, + 0.00434403121471405, + 0.006466447375714779, + 0.007950142957270145, + 0.009416354820132256, + 0.008944096975028515, + 0.009114086627960205, + 0.009963037446141243, + 0.007137936074286699, + 0.05994262546300888, + 0.018954217433929443, + 0.012660047970712185, + 0.01479420531541109, + 0.07316570729017258, + 0.17670437693595886, + 0.020758723840117455, + 0.16451364755630493, + 0.007659084629267454, + 0.014178561046719551, + 0.013339562341570854, + 0.016145115718245506, + 0.020266061648726463, + 0.015512996353209019, + 0.013095216825604439, + 0.01096826046705246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.0028186615090817213 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a life-size sculpture of a male discus thrower carved from white Carrara marble with gray veining?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle capturing a static shot of the sculpture on a rectangular pedestal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the statue animate by launching forward from the pedestal in a powerful leap and then hovering in the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue land with a thud, causing a cloud of dust to erupt and spiderweb-like cracks to radiate onto the polished floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a minimalist art gallery with high white walls, a dark gray concrete floor, and track lighting casting long 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": [ + -0.003614442190155387 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.003614442190155387, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614442190155387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09330473095178604, + 0.031852275133132935, + 0.023235954344272614, + 0.04177388548851013, + 0.0562136173248291, + 0.11813514679670334, + 0.09256070852279663, + 0.14468644559383392, + 0.36569198966026306, + 0.03254525735974312 + ], + "credit_map": [ + 0.0078257005661726, + 0.010953000746667385, + 0.009325305931270123, + 0.0161197017878294, + 0.013042786158621311, + 0.013183150440454483, + 0.010945233516395092, + 0.01190985832363367, + 0.002517939079552889, + 0.0035350569523870945, + 0.00251176580786705, + 0.005226833280175924, + 0.004178227391093969, + 0.0055054533295333385, + 0.0034256454091519117, + 0.004951353184878826, + 0.0016871494008228183, + 0.0025482468772679567, + 0.0013296684483066201, + 0.0032745732460170984, + 0.0034150765277445316, + 0.004537864122539759, + 0.0026173158548772335, + 0.003826058702543378, + 0.00208430178463459, + 0.0034614577889442444, + 0.0022764462046325207, + 0.0061361417174339294, + 0.0066708754748106, + 0.005913389381021261, + 0.008196764625608921, + 0.007034508511424065, + 0.0026633834931999445, + 0.006063490174710751, + 0.0031946233939379454, + 0.008648809045553207, + 0.007678603287786245, + 0.007765849586576223, + 0.012196294963359833, + 0.008002562448382378, + 0.0074029723182320595, + 0.01896681636571884, + 0.005561376456171274, + 0.012302862480282784, + 0.017862381413578987, + 0.017653118818998337, + 0.026911238208413124, + 0.011474376544356346, + 0.009007226675748825, + 0.0065366532653570175, + 0.015896353870630264, + 0.015764722600579262, + 0.005891732405871153, + 0.012970278039574623, + 0.014299836941063404, + 0.01219390518963337, + 0.014375755563378334, + 0.014613312669098377, + 0.014202572405338287, + 0.019445203244686127, + 0.017713496461510658, + 0.02124260738492012, + 0.024989547207951546, + 0.018103957176208496, + 0.040995169430971146, + 0.008705535903573036, + 0.007036696653813124, + 0.009206488728523254, + 0.050742194056510925, + 0.12388188391923904, + 0.012832775712013245, + 0.11229126900434494, + 0.002555612474679947, + 0.003273898735642433, + 0.002761124400421977, + 0.003707615192979574, + 0.00568058667704463, + 0.005061632487922907, + 0.004388001281768084, + 0.005116788670420647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.003614442190155387 + } + ], + "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/0000110/group_3/index_tensor([8430, 8430])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_tensor([8430, 8430])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..353468f9cc22143fd88a52cb1d28e06e5d440f52 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_tensor([8430, 8430])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium close-up shot, shot from a slightly elevated angle, focuses firmly on the center of a dilapidated ancient temple. A knight reclines against the pedestal of a massive, cracked stone Buddha with a compassionate countenance. His long, jet-black hair is casually tied back with a cloth ribbon, a few strands falling around his angular features. His chin bears the stubble of travel's weariness. He wears a frayed, dark blue, coarse cloth gown, the collar slightly open. His left hand rests seemingly casually on the hilt of an ancient, dark-brown leather-wrapped longsword. A cool, clear moonlight pours through a gaping hole in the temple's roof, creating a dramatic shaft of light. Countless motes of dust dance quietly within the beam, illuminating him and a corner of the Buddha statue behind him, creating a striking contrast between light and dark. In the deathly silence, his eyes, which had been tightly closed, suddenly opened. The whole movement was swift and decisive, without the slightest confusion or drowsiness. His pupils contracted instantly in the dim light, and his eyes were as sharp as an eagle, as if two substantial lightning bolts had cut through the stagnant air, carrying a cold murderous aura and absolute focus, shooting straight at an unknown existence outside the camera, instantly tearing the tranquility of the entire picture apart, and injecting extreme tension and an atmosphere of imminent danger.", + "seed": 13685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_8430_seed_13685_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": -0.03252950143069029, + "vlm_logprob": -0.03252950143069029, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03252950143069029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.015191215090453625 + ], + "opposite_logprob_tokens": [ + -4.515191078186035 + ], + "logprob_reward": -0.015191215090453625, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015191215090453625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48077869415283203, + 0.05131359025835991, + 0.03234107792377472, + 0.05690816044807434, + 0.05253610387444496, + 0.07538542151451111, + 0.053522661328315735, + 0.062137339264154434, + 0.10440192371606827, + 0.030675018206238747 + ], + "credit_map": [ + 0.17177467048168182, + 0.018991487100720406, + 0.01327076368033886, + 0.015484960749745369, + 0.016231782734394073, + 0.2057463675737381, + 0.012060101144015789, + 0.02721858210861683, + 0.0082538528367877, + 0.005912296939641237, + 0.0093950554728508, + 0.007675873581320047, + 0.002411843743175268, + 0.008894090540707111, + 0.003858487820252776, + 0.004912088625133038, + 0.005255234427750111, + 0.0034273373894393444, + 0.004125051200389862, + 0.007225459907203913, + 0.0017302652122452855, + 0.004106647800654173, + 0.002261576009914279, + 0.004209507256746292, + 0.009532362222671509, + 0.0064017982222139835, + 0.00793608371168375, + 0.007749665062874556, + 0.003700618864968419, + 0.01053793728351593, + 0.004811777733266354, + 0.006237920373678207, + 0.008632478304207325, + 0.005186921916902065, + 0.006812077946960926, + 0.011009753681719303, + 0.0030136783607304096, + 0.006596019491553307, + 0.005418611224740744, + 0.005866560153663158, + 0.011881314218044281, + 0.007318421732634306, + 0.008262681774795055, + 0.014195217750966549, + 0.0042170691303908825, + 0.0092203039675951, + 0.012891970574855804, + 0.007398441433906555, + 0.0067026433534920216, + 0.006727282889187336, + 0.011277923360466957, + 0.00496542826294899, + 0.005266073625534773, + 0.00840478390455246, + 0.005512407515197992, + 0.004666117951273918, + 0.00949524249881506, + 0.006740394048392773, + 0.01027037762105465, + 0.005822308361530304, + 0.006990781519562006, + 0.013504131697118282, + 0.004588632378727198, + 0.004725471138954163, + 0.02755490317940712, + 0.01017822790890932, + 0.011483852751553059, + 0.0096220588311553, + 0.010218534618616104, + 0.015495543368160725, + 0.005100989248603582, + 0.01474781148135662, + 0.005550864152610302, + 0.003830174682661891, + 0.002483103424310684, + 0.0023302563931792974, + 0.004593264311552048, + 0.005468924529850483, + 0.0019569401629269123, + 0.004461491014808416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.015191215090453625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.016079746186733246 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079746186733246, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079746186733246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2833341956138611, + 0.03728577867150307, + 0.0366412028670311, + 0.04425954446196556, + 0.08419571071863174, + 0.1962876170873642, + 0.09671220928430557, + 0.06764425337314606, + 0.12658095359802246, + 0.027058489620685577 + ], + "credit_map": [ + 0.09809792041778564, + 0.014089373871684074, + 0.0082607576623559, + 0.010603043250739574, + 0.009833340533077717, + 0.11753090471029282, + 0.008934811688959599, + 0.015984036028385162, + 0.004904885310679674, + 0.005783646833151579, + 0.004615715704858303, + 0.007300537079572678, + 0.0022515684831887484, + 0.005102386232465506, + 0.003617460373789072, + 0.003709575394168496, + 0.005689697805792093, + 0.0031150435097515583, + 0.006404552608728409, + 0.006142476573586464, + 0.00201943377032876, + 0.005541088525205851, + 0.00242390064522624, + 0.005305008497089148, + 0.008053337223827839, + 0.0054873754270374775, + 0.004927282687276602, + 0.006893265526741743, + 0.0022415712010115385, + 0.006627108436077833, + 0.004644110798835754, + 0.005385492462664843, + 0.015392729081213474, + 0.010388363152742386, + 0.010665114969015121, + 0.0141774145886302, + 0.003863003570586443, + 0.014140903949737549, + 0.006211535073816776, + 0.009356655180454254, + 0.029047034680843353, + 0.01672670803964138, + 0.02845608815550804, + 0.044889576733112335, + 0.017543792724609375, + 0.020591622218489647, + 0.0240696519613266, + 0.014963149093091488, + 0.012513542547821999, + 0.008106197230517864, + 0.027710288763046265, + 0.010092301294207573, + 0.007062732242047787, + 0.017072347924113274, + 0.00882387813180685, + 0.005330917425453663, + 0.011816148646175861, + 0.011001786217093468, + 0.01106932107359171, + 0.007113739848136902, + 0.0054168314673006535, + 0.01083665806800127, + 0.004733768757432699, + 0.005655996035784483, + 0.028014598414301872, + 0.011838104575872421, + 0.016423320397734642, + 0.011880084872245789, + 0.021872691810131073, + 0.01737687550485134, + 0.007059254217892885, + 0.012116032652556896, + 0.0043066018261015415, + 0.0032317033037543297, + 0.0029801849741488695, + 0.00267044547945261, + 0.0035822440404444933, + 0.004234527237713337, + 0.00211776839569211, + 0.003935013432055712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.016079746186733246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.03369637206196785 + ], + "opposite_logprob_tokens": [ + -3.408696413040161 + ], + "logprob_reward": -0.03369637206196785, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369637206196785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586463749408722, + 0.10651914775371552, + 0.04191436991095543, + 0.290092408657074, + 0.10478568077087402, + 0.044769734144210815, + 0.07496090978384018, + 0.06233447417616844, + 0.08083460479974747, + 0.03514227643609047 + ], + "credit_map": [ + 0.04376150667667389, + 0.010255783796310425, + 0.009624351747334003, + 0.015262169763445854, + 0.006709491368383169, + 0.052855078130960464, + 0.007287674583494663, + 0.012890326790511608, + 0.01770191267132759, + 0.010551716201007366, + 0.010000291280448437, + 0.025838952511548996, + 0.005359450820833445, + 0.014913156628608704, + 0.012429025024175644, + 0.009724635630846024, + 0.0063906325958669186, + 0.004681427963078022, + 0.0053993528708815575, + 0.006559919100254774, + 0.0023668664507567883, + 0.006662194617092609, + 0.003870696062222123, + 0.005983282346278429, + 0.04550870880484581, + 0.02640513889491558, + 0.03643878921866417, + 0.03612447530031204, + 0.01264527253806591, + 0.05979691445827484, + 0.02894521690905094, + 0.04422789439558983, + 0.018380658701062202, + 0.008252083323895931, + 0.012232590466737747, + 0.015460900031030178, + 0.005055761430412531, + 0.020520206540822983, + 0.007742307614535093, + 0.01714116893708706, + 0.007446294650435448, + 0.005520771257579327, + 0.005625245161354542, + 0.006158309057354927, + 0.0036710903514176607, + 0.00676044961437583, + 0.00476648285984993, + 0.004821090493351221, + 0.011917647905647755, + 0.009476401843130589, + 0.014258990995585918, + 0.007679634727537632, + 0.0066144634038209915, + 0.011640649288892746, + 0.007559845224022865, + 0.005813272669911385, + 0.010696614161133766, + 0.005339330993592739, + 0.010443845763802528, + 0.004724040161818266, + 0.005241405218839645, + 0.014891330152750015, + 0.006125043611973524, + 0.004872864577919245, + 0.015388119034469128, + 0.008458160795271397, + 0.009619132615625858, + 0.008125007152557373, + 0.015556947328150272, + 0.010013609193265438, + 0.005350905004888773, + 0.008322724141180515, + 0.004621976520866156, + 0.0037083772476762533, + 0.0028714709915220737, + 0.0031664129346609116, + 0.006905943620949984, + 0.00552402762696147, + 0.0036640409380197525, + 0.004680026788264513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.03369637206196785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.008637331426143646 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637331426143646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637331426143646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29031574726104736, + 0.04476993530988693, + 0.039775725454092026, + 0.10521005094051361, + 0.06371703743934631, + 0.13093814253807068, + 0.06624315679073334, + 0.08431434631347656, + 0.12141772359609604, + 0.05329814925789833 + ], + "credit_map": [ + 0.099873386323452, + 0.014114150777459145, + 0.009602649137377739, + 0.012804566882550716, + 0.010423734784126282, + 0.11805006861686707, + 0.008033049292862415, + 0.017414139583706856, + 0.006802862044423819, + 0.005716245621442795, + 0.00908140279352665, + 0.005469653755426407, + 0.0024440435227006674, + 0.006897172890603542, + 0.0033662309870123863, + 0.004992324393242598, + 0.005960316862910986, + 0.003989106975495815, + 0.004006023984402418, + 0.005152217112481594, + 0.002993164351209998, + 0.007583870552480221, + 0.0033215780276805162, + 0.006769447587430477, + 0.017768001183867455, + 0.011142781935632229, + 0.014886947348713875, + 0.01314529962837696, + 0.004863308276981115, + 0.02001437544822693, + 0.009139410220086575, + 0.014249925501644611, + 0.009743972681462765, + 0.004864515271037817, + 0.008155963383615017, + 0.006025186739861965, + 0.011487441137433052, + 0.0122957369312644, + 0.0040318830870091915, + 0.007112343329936266, + 0.021160289645195007, + 0.012006744742393494, + 0.016045089811086655, + 0.02235143631696701, + 0.007352503016591072, + 0.018543969839811325, + 0.01976640708744526, + 0.013711688108742237, + 0.008509598672389984, + 0.009333614259958267, + 0.011897730641067028, + 0.006752450484782457, + 0.005895428359508514, + 0.009446371346712112, + 0.006639579776674509, + 0.00776838231831789, + 0.011264866217970848, + 0.006221882998943329, + 0.013588142581284046, + 0.005807684268802404, + 0.009637455455958843, + 0.01819228194653988, + 0.00965260248631239, + 0.009949435479938984, + 0.016230730339884758, + 0.013894307427108288, + 0.012072502635419369, + 0.011847248300909996, + 0.025898024439811707, + 0.016205331310629845, + 0.011448155157268047, + 0.013821425847709179, + 0.004464276134967804, + 0.004373482428491116, + 0.004835322964936495, + 0.003875449299812317, + 0.010846586897969246, + 0.008235504850745201, + 0.006969587877392769, + 0.009697943925857544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.008637331426143646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.08904284238815308 + ], + "opposite_logprob_tokens": [ + -2.464042901992798 + ], + "logprob_reward": -0.08904284238815308, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904284238815308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16837070882320404, + 0.041809987276792526, + 0.030673271045088768, + 0.04703431576490402, + 0.06977587938308716, + 0.05863814428448677, + 0.1498592048883438, + 0.17454855144023895, + 0.2314748615026474, + 0.027815107256174088 + ], + "credit_map": [ + 0.05110843852162361, + 0.010160310193896294, + 0.011131355538964272, + 0.011084459722042084, + 0.006135350093245506, + 0.059687577188014984, + 0.008410985581576824, + 0.010652218945324421, + 0.005152315832674503, + 0.006497056223452091, + 0.006057555787265301, + 0.005618126131594181, + 0.0024687584955245256, + 0.005203562323004007, + 0.00608165143057704, + 0.004730957094579935, + 0.004947553854435682, + 0.002580763306468725, + 0.003548300825059414, + 0.0030635837465524673, + 0.0018359236419200897, + 0.006660746410489082, + 0.0022124850656837225, + 0.005823912564665079, + 0.00768428947776556, + 0.005952587816864252, + 0.009478282183408737, + 0.008249175734817982, + 0.0017950359033420682, + 0.005904343444854021, + 0.0030156702268868685, + 0.004954925738275051, + 0.012349712662398815, + 0.007283895742148161, + 0.008684050291776657, + 0.006934456992894411, + 0.0032587721943855286, + 0.013312346301972866, + 0.004284691531211138, + 0.013667949475347996, + 0.009361973963677883, + 0.006464607082307339, + 0.007293994072824717, + 0.009795432910323143, + 0.006913107354193926, + 0.006938734091818333, + 0.007503433618694544, + 0.0043668607249855995, + 0.0189460888504982, + 0.014093323610723019, + 0.02882082387804985, + 0.01787300780415535, + 0.011443577706813812, + 0.021829532459378242, + 0.017474357038736343, + 0.019378473982214928, + 0.028479302302002907, + 0.013363860547542572, + 0.03491923585534096, + 0.00935627706348896, + 0.01300118863582611, + 0.04405627399682999, + 0.01693841442465782, + 0.014433996751904488, + 0.04658058285713196, + 0.02672983519732952, + 0.02788846381008625, + 0.02483413554728031, + 0.03595218062400818, + 0.03078768216073513, + 0.013828729279339314, + 0.02487323433160782, + 0.004558219108730555, + 0.0033090694341808558, + 0.0024788803420960903, + 0.002865449758246541, + 0.004464050754904747, + 0.0041987095028162, + 0.002324839122593403, + 0.003615886438637972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.08904284238815308 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot, shot from a slightly elevated angle, focuses firmly on the center of a dilapidated ancient temple. A knight reclines against the pedestal of a massive, cracked stone Buddha with a compassionate countenance. His long, jet-black hair is casually tied back with a cloth ribbon, a few strands falling around his angular features. His chin bears the stubble of travel's weariness. He wears a frayed, dark blue, coarse cloth gown, the collar slightly open. His left hand rests seemingly casually on the hilt of an ancient, dark-brown leather-wrapped longsword. A cool, clear moonlight pours through a gaping hole in the temple's roof, creating a dramatic shaft of light. Countless motes of dust dance quietly within the beam, illuminating him and a corner of the Buddha statue behind him, creating a striking contrast between light and dark. In the deathly silence, his eyes, which had been tightly closed, suddenly opened. The whole movement was swift and decisive, without the slightest confusion or drowsiness. His pupils contracted instantly in the dim light, and his eyes were as sharp as an eagle, as if two substantial lightning bolts had cut through the stagnant air, carrying a cold murderous aura and absolute focus, shooting straight at an unknown existence outside the camera, instantly tearing the tranquility of the entire picture apart, and injecting extreme tension and an atmosphere of imminent danger.", + "seed": 13686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_8430_seed_13686_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": -0.026694660354405642, + "vlm_logprob": -0.026694660354405642, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.026694660354405642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.015198964625597 + ], + "opposite_logprob_tokens": [ + -4.515199184417725 + ], + "logprob_reward": -0.015198964625597, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015198964625597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035819411277770996, + 0.0286323893815279, + 0.02841312065720558, + 0.030126171186566353, + 0.03194492310285568, + 0.16207431256771088, + 0.5510117411613464, + 0.04295370355248451, + 0.06306871771812439, + 0.02595548890531063 + ], + "credit_map": [ + 0.0041369665414094925, + 0.004753689281642437, + 0.003666527569293976, + 0.004169232677668333, + 0.00552526768296957, + 0.004810125567018986, + 0.0041056242771446705, + 0.004651972558349371, + 0.0038129198364913464, + 0.004584738984704018, + 0.0029291885439306498, + 0.003245010506361723, + 0.003615873632952571, + 0.0032017389312386513, + 0.0035377745516598225, + 0.003705143928527832, + 0.0032509551383554935, + 0.0024999172892421484, + 0.003115972736850381, + 0.0030857757665216923, + 0.005017440300434828, + 0.002975690644234419, + 0.004686153028160334, + 0.0037812148220837116, + 0.004718294367194176, + 0.003651607548817992, + 0.003505895147100091, + 0.0024989573284983635, + 0.003629646496847272, + 0.003988184034824371, + 0.005020475015044212, + 0.0031131107825785875, + 0.003196772187948227, + 0.00461886590346694, + 0.0026790096890181303, + 0.0033844211138784885, + 0.0043793912045657635, + 0.0039397492073476315, + 0.00462676165625453, + 0.005119947250932455, + 0.013575390912592411, + 0.019130166620016098, + 0.014363625086843967, + 0.021581899374723434, + 0.03536490350961685, + 0.01909996010363102, + 0.021230492740869522, + 0.017727861180901527, + 0.025866655632853508, + 0.015934908762574196, + 0.054631590843200684, + 0.015854327008128166, + 0.014890490099787712, + 0.22874878346920013, + 0.15400297939777374, + 0.04108201339840889, + 0.006956006400287151, + 0.00628905463963747, + 0.0055787209421396255, + 0.0030831454787403345, + 0.004616112448275089, + 0.0051698521710932255, + 0.0061834026128053665, + 0.005077405832707882, + 0.006146230734884739, + 0.015062550082802773, + 0.00581591110676527, + 0.009757242165505886, + 0.004804579541087151, + 0.008075528778135777, + 0.007397567853331566, + 0.006009101402014494, + 0.0028798214625567198, + 0.004933306016027927, + 0.003045513294637203, + 0.00308470381423831, + 0.001379560213536024, + 0.0037873457185924053, + 0.0036206368822604418, + 0.0032246001064777374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.015198964625597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.020596787333488464 + ], + "opposite_logprob_tokens": [ + -3.895596742630005 + ], + "logprob_reward": -0.020596787333488464, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596787333488464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07359843701124191, + 0.04064241051673889, + 0.08644749969244003, + 0.07295837998390198, + 0.04173455387353897, + 0.0708104819059372, + 0.09234607219696045, + 0.18190889060497284, + 0.16204586625099182, + 0.1775074303150177 + ], + "credit_map": [ + 0.011776396073400974, + 0.009105096571147442, + 0.007394730579108, + 0.009110295213758945, + 0.008653593249619007, + 0.009198877960443497, + 0.009646710008382797, + 0.008712735958397388, + 0.006696338299661875, + 0.006021264009177685, + 0.003120432374998927, + 0.0045265816152095795, + 0.004814059007912874, + 0.0047357771545648575, + 0.005339060910046101, + 0.005388896446675062, + 0.008328012190759182, + 0.006483871955424547, + 0.007548120804131031, + 0.008787494152784348, + 0.014982140623033047, + 0.008892187848687172, + 0.019198844209313393, + 0.012226822786033154, + 0.009947285056114197, + 0.007799863815307617, + 0.00847203005105257, + 0.008851156570017338, + 0.009089905768632889, + 0.008378978818655014, + 0.011544040404260159, + 0.008875119499862194, + 0.005250851158052683, + 0.004524271935224533, + 0.00535140885040164, + 0.005629158578813076, + 0.0068959989584982395, + 0.004279616288840771, + 0.005572732072323561, + 0.004230516962707043, + 0.008490865118801594, + 0.0056165894493460655, + 0.007850012741982937, + 0.0053742132149636745, + 0.011722082272171974, + 0.008917033672332764, + 0.011732528917491436, + 0.011107156984508038, + 0.007181691005825996, + 0.0066358852200210094, + 0.018974417820572853, + 0.005178005900233984, + 0.005013477057218552, + 0.02579156868159771, + 0.009333673864603043, + 0.014237351715564728, + 0.01991666853427887, + 0.030769383534789085, + 0.016617219895124435, + 0.014242173172533512, + 0.0323667973279953, + 0.020809663459658623, + 0.027398576959967613, + 0.019788414239883423, + 0.014071370474994183, + 0.019308749586343765, + 0.019108373671770096, + 0.019495176151394844, + 0.021095246076583862, + 0.021013619378209114, + 0.029367517679929733, + 0.018585817888379097, + 0.019014406949281693, + 0.0256081260740757, + 0.024726081639528275, + 0.02017456479370594, + 0.022800978273153305, + 0.02252708375453949, + 0.023287562653422356, + 0.019368620589375496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.020596787333488464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.04306785389780998 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.04306785389780998, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306785389780998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15343397855758667, + 0.07801323384046555, + 0.07973919063806534, + 0.07191561907529831, + 0.04715092107653618, + 0.11767134815454483, + 0.25597119331359863, + 0.06865570694208145, + 0.07875755429267883, + 0.048691269010305405 + ], + "credit_map": [ + 0.01583494432270527, + 0.023421507328748703, + 0.013002931140363216, + 0.0232998076826334, + 0.017494039610028267, + 0.019114507362246513, + 0.02593952603638172, + 0.015326698310673237, + 0.008263543248176575, + 0.013194001279771328, + 0.010450396686792374, + 0.010382312349975109, + 0.009138185530900955, + 0.0076671079732477665, + 0.007755402009934187, + 0.011162289418280125, + 0.00918402150273323, + 0.009563710540533066, + 0.00675223721191287, + 0.01091943122446537, + 0.012516629882156849, + 0.009545713663101196, + 0.014159153215587139, + 0.007098297588527203, + 0.009599190205335617, + 0.008219819515943527, + 0.007531961426138878, + 0.008417497389018536, + 0.011114517226815224, + 0.007596935145556927, + 0.011607805266976357, + 0.007827895693480968, + 0.005486139561980963, + 0.005772274918854237, + 0.008335010148584843, + 0.005455782171338797, + 0.005267094820737839, + 0.005915748421102762, + 0.006810361053794622, + 0.004108508117496967, + 0.010979875922203064, + 0.012587360106408596, + 0.011450625956058502, + 0.011097813956439495, + 0.02009371481835842, + 0.01892208494246006, + 0.014026664197444916, + 0.018513204529881477, + 0.01175251416862011, + 0.00844112690538168, + 0.02072412706911564, + 0.01229894533753395, + 0.00981492456048727, + 0.1011359691619873, + 0.07135668396949768, + 0.02044691890478134, + 0.006643247790634632, + 0.008645850233733654, + 0.006627518218010664, + 0.010220153257250786, + 0.006442549172788858, + 0.008631706237792969, + 0.010820388793945312, + 0.01062429416924715, + 0.0070357476361095905, + 0.01332341879606247, + 0.009817151352763176, + 0.013970223255455494, + 0.004358682315796614, + 0.011029921472072601, + 0.011549805290997028, + 0.00767260417342186, + 0.004824230447411537, + 0.00832116138190031, + 0.004704622086137533, + 0.007938091643154621, + 0.0025470659602433443, + 0.006351379211992025, + 0.008947104215621948, + 0.005057611968368292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.04306785389780998 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2431633472442627, + 0.04140298813581467, + 0.12106433510780334, + 0.06097130849957466, + 0.11398523300886154, + 0.04494577646255493, + 0.16811883449554443, + 0.05233755707740784, + 0.08668313175439835, + 0.06732743233442307 + ], + "credit_map": [ + 0.038723599165678024, + 0.028385955840349197, + 0.018940461799502373, + 0.01941506564617157, + 0.048545412719249725, + 0.03265431523323059, + 0.03253868967294693, + 0.023959875106811523, + 0.006190779618918896, + 0.00711971428245306, + 0.0026170541532337666, + 0.003282182849943638, + 0.006181178614497185, + 0.00465491833165288, + 0.004490381106734276, + 0.006866783834993839, + 0.0159764364361763, + 0.00829449575394392, + 0.010426284745335579, + 0.01190239004790783, + 0.01701761968433857, + 0.013972830958664417, + 0.028334900736808777, + 0.015139388851821423, + 0.011547069996595383, + 0.007186849135905504, + 0.00666007399559021, + 0.007012091111391783, + 0.006977519486099482, + 0.007943647913634777, + 0.0068517085164785385, + 0.006792351603507996, + 0.018973898142576218, + 0.012173437513411045, + 0.013381258584558964, + 0.011089364998042583, + 0.0116061270236969, + 0.015282454900443554, + 0.015580443665385246, + 0.01589825376868248, + 0.0037627725396305323, + 0.005720381624996662, + 0.003225669264793396, + 0.005692858714610338, + 0.0063161058351397514, + 0.005538146942853928, + 0.006064382381737232, + 0.008625464513897896, + 0.008784222416579723, + 0.004788189195096493, + 0.009936494752764702, + 0.0069461301900446415, + 0.005645002704113722, + 0.06721223890781403, + 0.050836268812417984, + 0.01397030521184206, + 0.004792093299329281, + 0.007052266038954258, + 0.004108953755348921, + 0.005539525765925646, + 0.005943187978118658, + 0.0063649872317910194, + 0.010277371853590012, + 0.008259176276624203, + 0.0068176379427313805, + 0.013238140381872654, + 0.01114541944116354, + 0.012917395681142807, + 0.006276674102991819, + 0.012079600244760513, + 0.014532875269651413, + 0.009675397537648678, + 0.007890546694397926, + 0.008125566877424717, + 0.012094473466277122, + 0.007399675901979208, + 0.004208043217658997, + 0.009129950776696205, + 0.009330986998975277, + 0.009148192591965199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.04865886643528938 + ], + "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": [ + 0.06899593025445938, + 0.05399448052048683, + 0.046556469053030014, + 0.06463444977998734, + 0.06271596252918243, + 0.08287729322910309, + 0.27111542224884033, + 0.14135242998600006, + 0.11445077508687973, + 0.0933067724108696 + ], + "credit_map": [ + 0.008542004972696304, + 0.008326920680701733, + 0.007388444617390633, + 0.009905175305902958, + 0.00829579308629036, + 0.00881657749414444, + 0.010065858252346516, + 0.007655153516680002, + 0.008600938133895397, + 0.008156646974384785, + 0.003912551794201136, + 0.006060146726667881, + 0.006867383141070604, + 0.0067991879768669605, + 0.00707817729562521, + 0.0065194484777748585, + 0.005102938041090965, + 0.0037462327163666487, + 0.004797317553311586, + 0.003741899970918894, + 0.007221609354019165, + 0.00537052471190691, + 0.010511772707104683, + 0.006064177490770817, + 0.007859953679144382, + 0.0069211507216095924, + 0.006062958389520645, + 0.004350457340478897, + 0.010683294385671616, + 0.007775591220706701, + 0.012701323255896568, + 0.008279724977910519, + 0.006996958050876856, + 0.007526993285864592, + 0.006902238354086876, + 0.009439787827432156, + 0.009343753568828106, + 0.007971236482262611, + 0.007016960997134447, + 0.007518029306083918, + 0.008902876637876034, + 0.00673517445102334, + 0.007979735732078552, + 0.008660288527607918, + 0.014253858476877213, + 0.011572285555303097, + 0.012962952256202698, + 0.011810126714408398, + 0.015627160668373108, + 0.01791548542678356, + 0.02387274242937565, + 0.012103079818189144, + 0.016210416331887245, + 0.09246622771024704, + 0.06919746845960617, + 0.023722851648926735, + 0.013113990426063538, + 0.021738538518548012, + 0.00960309337824583, + 0.013754351995885372, + 0.016195476055145264, + 0.01705898903310299, + 0.030080627650022507, + 0.0198073647916317, + 0.010972008109092712, + 0.017250055447220802, + 0.012789892964065075, + 0.021162668243050575, + 0.005876181181520224, + 0.015756061300635338, + 0.018614523112773895, + 0.012029388919472694, + 0.008963122963905334, + 0.019229386001825333, + 0.009334398433566093, + 0.018569914624094963, + 0.007144031580537558, + 0.011188795790076256, + 0.010532855987548828, + 0.008344265632331371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.04865886643528938 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot, shot from a slightly elevated angle, focuses firmly on the center of a dilapidated ancient temple. A knight reclines against the pedestal of a massive, cracked stone Buddha with a compassionate countenance. His long, jet-black hair is casually tied back with a cloth ribbon, a few strands falling around his angular features. His chin bears the stubble of travel's weariness. He wears a frayed, dark blue, coarse cloth gown, the collar slightly open. His left hand rests seemingly casually on the hilt of an ancient, dark-brown leather-wrapped longsword. A cool, clear moonlight pours through a gaping hole in the temple's roof, creating a dramatic shaft of light. Countless motes of dust dance quietly within the beam, illuminating him and a corner of the Buddha statue behind him, creating a striking contrast between light and dark. In the deathly silence, his eyes, which had been tightly closed, suddenly opened. The whole movement was swift and decisive, without the slightest confusion or drowsiness. His pupils contracted instantly in the dim light, and his eyes were as sharp as an eagle, as if two substantial lightning bolts had cut through the stagnant air, carrying a cold murderous aura and absolute focus, shooting straight at an unknown existence outside the camera, instantly tearing the tranquility of the entire picture apart, and injecting extreme tension and an atmosphere of imminent danger.", + "seed": 13687, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_8430_seed_13687_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": -0.020105081237852575, + "vlm_logprob": -0.020105081237852575, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.020105081237852575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.00958495493978262 + ], + "opposite_logprob_tokens": [ + -4.884584903717041 + ], + "logprob_reward": -0.00958495493978262, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00958495493978262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0966171845793724, + 0.0500367134809494, + 0.03820990398526192, + 0.034456826746463776, + 0.04280918464064598, + 0.06213631480932236, + 0.08907295018434525, + 0.3903557062149048, + 0.08145222067832947, + 0.11485292762517929 + ], + "credit_map": [ + 0.013765575364232063, + 0.024813037365674973, + 0.012559142895042896, + 0.010685907676815987, + 0.009987504221498966, + 0.0075509026646614075, + 0.007297446019947529, + 0.009957674890756607, + 0.009881026111543179, + 0.006995106115937233, + 0.012195550836622715, + 0.00426337867975235, + 0.003337630070745945, + 0.004740743897855282, + 0.004763781558722258, + 0.0038594952784478664, + 0.0061333454214036465, + 0.005413016304373741, + 0.007150501944124699, + 0.003486348083242774, + 0.003546815598383546, + 0.004824312869459391, + 0.004141605459153652, + 0.00351395970210433, + 0.004832556936889887, + 0.006476050708442926, + 0.00338984583504498, + 0.004982029553502798, + 0.0036266224924474955, + 0.0029694580007344484, + 0.004558122716844082, + 0.003622140269726515, + 0.005759967025369406, + 0.011090910993516445, + 0.00538887782022357, + 0.004933180287480354, + 0.003923723474144936, + 0.0024550368543714285, + 0.00438534514978528, + 0.004872145131230354, + 0.015946311876177788, + 0.010290471836924553, + 0.006512471940368414, + 0.006610159762203693, + 0.006230473518371582, + 0.005841044709086418, + 0.004819923080503941, + 0.005885463673621416, + 0.00746350921690464, + 0.027857942506670952, + 0.009089821949601173, + 0.008491506800055504, + 0.008776471018791199, + 0.006265982985496521, + 0.012969226576387882, + 0.008158492855727673, + 0.03570122271776199, + 0.025465138256549835, + 0.017439894378185272, + 0.02387065440416336, + 0.038293931633234024, + 0.027171436697244644, + 0.10399157553911209, + 0.11842186748981476, + 0.008807516656816006, + 0.012524877674877644, + 0.007392249070107937, + 0.00509460037574172, + 0.014896969310939312, + 0.006556726526468992, + 0.012146077118813992, + 0.014033207669854164, + 0.024077853187918663, + 0.02182883769273758, + 0.011998211964964867, + 0.014541401527822018, + 0.01590142212808132, + 0.008370449766516685, + 0.011599119752645493, + 0.006535633932799101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.00958495493978262 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.016081271693110466 + ], + "opposite_logprob_tokens": [ + -4.1410813331604 + ], + "logprob_reward": -0.016081271693110466, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081271693110466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05880780145525932, + 0.03447901830077171, + 0.04851263761520386, + 0.03713707998394966, + 0.04336148500442505, + 0.06209128350019455, + 0.11978880316019058, + 0.43594616651535034, + 0.06811480224132538, + 0.09176082164049149 + ], + "credit_map": [ + 0.007596759125590324, + 0.013133969157934189, + 0.008955228142440319, + 0.007441342808306217, + 0.006762185133993626, + 0.004123871214687824, + 0.0050291018560528755, + 0.0057653565891087055, + 0.00525248097255826, + 0.008255891501903534, + 0.005187168251723051, + 0.0038499836809933186, + 0.001977450679987669, + 0.0018964038463309407, + 0.004248206503689289, + 0.003811439499258995, + 0.007406304590404034, + 0.0057532330974936485, + 0.0073206378147006035, + 0.006163685582578182, + 0.006358340848237276, + 0.004123524762690067, + 0.005217758007347584, + 0.00616916548460722, + 0.005063768941909075, + 0.005143139511346817, + 0.004917447455227375, + 0.00419864896684885, + 0.00481004361063242, + 0.0029556783847510815, + 0.005988900549709797, + 0.004059457220137119, + 0.0060172490775585175, + 0.007173598278313875, + 0.005448644980788231, + 0.004396253731101751, + 0.005591500550508499, + 0.0031382874585688114, + 0.004917108453810215, + 0.00667885085567832, + 0.013592901639640331, + 0.009441699832677841, + 0.00541809992864728, + 0.004435691051185131, + 0.012267584912478924, + 0.006081258412450552, + 0.006106653716415167, + 0.00474740331992507, + 0.012391484342515469, + 0.03245749697089195, + 0.009819786995649338, + 0.01442316360771656, + 0.009625867940485477, + 0.00750127574428916, + 0.01328530814498663, + 0.020284445956349373, + 0.04230910912156105, + 0.0874238982796669, + 0.019346073269844055, + 0.02105785347521305, + 0.02461901120841503, + 0.028461331501603127, + 0.07101329416036606, + 0.1417156457901001, + 0.0068419743329286575, + 0.00917147658765316, + 0.007029019296169281, + 0.004293118137866259, + 0.012335268780589104, + 0.004594980739057064, + 0.014903411269187927, + 0.00894556101411581, + 0.015315092168748379, + 0.015930304303765297, + 0.010558176785707474, + 0.008200663141906261, + 0.016056526452302933, + 0.008399847894906998, + 0.011371690779924393, + 0.005928535480052233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.016081271693110466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.0336916446685791 + ], + "opposite_logprob_tokens": [ + -3.408691644668579 + ], + "logprob_reward": -0.0336916446685791, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336916446685791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12771889567375183, + 0.04417908936738968, + 0.02933095023036003, + 0.023951880633831024, + 0.026569068431854248, + 0.028298132121562958, + 0.02968117594718933, + 0.5894305109977722, + 0.0470375120639801, + 0.053802717477083206 + ], + "credit_map": [ + 0.015485490672290325, + 0.015810521319508553, + 0.015824047848582268, + 0.020146381109952927, + 0.0112860556691885, + 0.010791945271193981, + 0.01963733322918415, + 0.018737133592367172, + 0.006531998980790377, + 0.006591195706278086, + 0.007725220173597336, + 0.005908837076276541, + 0.003235599957406521, + 0.002994819078594446, + 0.005736680701375008, + 0.005454742815345526, + 0.005135394167155027, + 0.0038329234812408686, + 0.005694519262760878, + 0.002838424639776349, + 0.0028469176031649113, + 0.003175352467224002, + 0.0028414493426680565, + 0.002965973224490881, + 0.0033702279906719923, + 0.003337965114042163, + 0.003017607843503356, + 0.004429536405950785, + 0.002538054483011365, + 0.0017156768590211868, + 0.002724002581089735, + 0.0028188105206936598, + 0.003934973850846291, + 0.006580338813364506, + 0.0035711620002985, + 0.0033449947368353605, + 0.0025648297742009163, + 0.0019770795479416847, + 0.002325270092114806, + 0.002270422177389264, + 0.003993657883256674, + 0.004389884416013956, + 0.0028887970838695765, + 0.0020465233828872442, + 0.005747586488723755, + 0.003644210984930396, + 0.0030216325540095568, + 0.002565844217315316, + 0.005578434094786644, + 0.004318397957831621, + 0.0027824605349451303, + 0.0028403440956026316, + 0.0050761825405061245, + 0.002034256001934409, + 0.003651995211839676, + 0.00339910970069468, + 0.05411769822239876, + 0.05392728000879288, + 0.023592744022607803, + 0.03699534386396408, + 0.05420869588851929, + 0.037147533148527145, + 0.17337152361869812, + 0.15606974065303802, + 0.005772152915596962, + 0.009700010530650616, + 0.005364740267395973, + 0.0032689333893358707, + 0.008044067770242691, + 0.0025310309138149023, + 0.007112232502549887, + 0.005244348663836718, + 0.008508387953042984, + 0.007686269469559193, + 0.006299220025539398, + 0.00417785719037056, + 0.009803050197660923, + 0.004105876199901104, + 0.006791849620640278, + 0.006430213805288076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0336916446685791 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.02058207429945469 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.02058207429945469, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058207429945469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1351509541273117, + 0.059736624360084534, + 0.05974334478378296, + 0.023561015725135803, + 0.0750790610909462, + 0.043128952383995056, + 0.12843964993953705, + 0.2154451310634613, + 0.10734561830759048, + 0.1523696482181549 + ], + "credit_map": [ + 0.02178625576198101, + 0.02538587525486946, + 0.023786261677742004, + 0.014907537959516048, + 0.010781988501548767, + 0.01153957936912775, + 0.011866927146911621, + 0.015096521936357021, + 0.012048548087477684, + 0.0101866340264678, + 0.014206742867827415, + 0.006185619160532951, + 0.0025435821153223515, + 0.0045047784224152565, + 0.005056951195001602, + 0.005003770347684622, + 0.010985658504068851, + 0.013891499489545822, + 0.012414058670401573, + 0.004798016045242548, + 0.0034745503216981888, + 0.0063674235716462135, + 0.0034747400786727667, + 0.0043373978696763515, + 0.0038473461754620075, + 0.0036676842719316483, + 0.004012148827314377, + 0.0026904987171292305, + 0.0024264228995889425, + 0.002429765649139881, + 0.0022084233351051807, + 0.002278727013617754, + 0.012480512261390686, + 0.023547658696770668, + 0.010497107170522213, + 0.006469722371548414, + 0.005594115704298019, + 0.005629055202007294, + 0.004679660778492689, + 0.0061812386848032475, + 0.013170870952308178, + 0.008682524785399437, + 0.004513629712164402, + 0.0032910052686929703, + 0.0049736774526536465, + 0.0031288829632103443, + 0.002804963383823633, + 0.002563399262726307, + 0.015886511653661728, + 0.044993992894887924, + 0.01948140375316143, + 0.009709188714623451, + 0.008713250048458576, + 0.006107150111347437, + 0.018414627760648727, + 0.005133519414812326, + 0.019189409911632538, + 0.01675543375313282, + 0.008638909086585045, + 0.013630029745399952, + 0.020063186064362526, + 0.015149185433983803, + 0.061145931482315063, + 0.06087305396795273, + 0.02001887932419777, + 0.026215609163045883, + 0.015903491526842117, + 0.009816541336476803, + 0.007617277558892965, + 0.005884195677936077, + 0.012187041342258453, + 0.009702587500214577, + 0.04411431774497032, + 0.03427639976143837, + 0.01549791730940342, + 0.01942598633468151, + 0.009579299949109554, + 0.009108491241931915, + 0.011647814884781837, + 0.008719426579773426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.02058207429945469 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048320069909095764, + 0.032548241317272186, + 0.020115619525313377, + 0.02523217350244522, + 0.02030949853360653, + 0.04163513705134392, + 0.10898391157388687, + 0.271562397480011, + 0.07626331597566605, + 0.3550296425819397 + ], + "credit_map": [ + 0.006909408140927553, + 0.008603198453783989, + 0.006617013365030289, + 0.006393219344317913, + 0.003416278399527073, + 0.0044228374026715755, + 0.0056799850426614285, + 0.006278129760175943, + 0.004352339077740908, + 0.006529208272695541, + 0.00549355847761035, + 0.004505733959376812, + 0.0020503117702901363, + 0.0024511581286787987, + 0.003993619699031115, + 0.003172310534864664, + 0.003181030508130789, + 0.003836055751889944, + 0.0035354350693523884, + 0.002124067395925522, + 0.0018073326209560037, + 0.0016965154791250825, + 0.0018712175078690052, + 0.002063965192064643, + 0.003727163653820753, + 0.004140689503401518, + 0.0029128699097782373, + 0.002605166519060731, + 0.0034463151823729277, + 0.0020862696692347527, + 0.0035063051618635654, + 0.002807393902912736, + 0.0032544576097279787, + 0.004142999183386564, + 0.002112820977345109, + 0.0022167887073010206, + 0.0020841415971517563, + 0.0016782396705821157, + 0.002320480067282915, + 0.002499569207429886, + 0.004515530075877905, + 0.006512691266834736, + 0.004693173337727785, + 0.003095735562965274, + 0.006287472788244486, + 0.005884707905352116, + 0.004133086651563644, + 0.006512738298624754, + 0.012405592948198318, + 0.03294000402092934, + 0.011064788326621056, + 0.01031547226011753, + 0.00783043634146452, + 0.007115878164768219, + 0.010314828716218472, + 0.01699690707027912, + 0.0272530447691679, + 0.05979527160525322, + 0.016785837709903717, + 0.018809927627444267, + 0.02829825133085251, + 0.01178373396396637, + 0.07826469838619232, + 0.030571602284908295, + 0.008187747560441494, + 0.008485662750899792, + 0.006490940693765879, + 0.005848018452525139, + 0.012525385245680809, + 0.006502394564449787, + 0.01653134822845459, + 0.011691821739077568, + 0.0811806321144104, + 0.06526565551757812, + 0.04280878230929375, + 0.03686249256134033, + 0.04132977873086929, + 0.029692893847823143, + 0.03148750960826874, + 0.026401899755001068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.02058546058833599 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up shot, shot from a slightly elevated angle, focuses firmly on the center of a dilapidated ancient temple. A knight reclines against the pedestal of a massive, cracked stone Buddha with a compassionate countenance. His long, jet-black hair is casually tied back with a cloth ribbon, a few strands falling around his angular features. His chin bears the stubble of travel's weariness. He wears a frayed, dark blue, coarse cloth gown, the collar slightly open. His left hand rests seemingly casually on the hilt of an ancient, dark-brown leather-wrapped longsword. A cool, clear moonlight pours through a gaping hole in the temple's roof, creating a dramatic shaft of light. Countless motes of dust dance quietly within the beam, illuminating him and a corner of the Buddha statue behind him, creating a striking contrast between light and dark. In the deathly silence, his eyes, which had been tightly closed, suddenly opened. The whole movement was swift and decisive, without the slightest confusion or drowsiness. His pupils contracted instantly in the dim light, and his eyes were as sharp as an eagle, as if two substantial lightning bolts had cut through the stagnant air, carrying a cold murderous aura and absolute focus, shooting straight at an unknown existence outside the camera, instantly tearing the tranquility of the entire picture apart, and injecting extreme tension and an atmosphere of imminent danger.", + "seed": 13688, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_3/index_8430_seed_13688_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": -0.02009393656626344, + "vlm_logprob": -0.02009393656626344, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02009393656626344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.018498403951525688 + ], + "opposite_logprob_tokens": [ + -4.393498420715332 + ], + "logprob_reward": -0.018498403951525688, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.018498403951525688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023555535823106766, + 0.010980038903653622, + 0.011302648112177849, + 0.012542500160634518, + 0.0227219108492136, + 0.030360745266079903, + 0.051659952849149704, + 0.27883660793304443, + 0.547157883644104, + 0.010882225818932056 + ], + "credit_map": [ + 0.002549160039052367, + 0.0021887561306357384, + 0.003629234852269292, + 0.002814567182213068, + 0.0028194645419716835, + 0.0037678105290979147, + 0.002760554663836956, + 0.003025986487045884, + 0.001491811708547175, + 0.0006278135697357357, + 0.0009532828698866069, + 0.0017350023845210671, + 0.0016696879174560308, + 0.0015792190097272396, + 0.001518096774816513, + 0.0014051252510398626, + 0.0014210119843482971, + 0.0007882725913077593, + 0.0015499363653361797, + 0.0013179488014429808, + 0.0014447750290855765, + 0.0016877169255167246, + 0.0016319608548656106, + 0.001461024396121502, + 0.001751248026266694, + 0.0010927369585260749, + 0.0018394049257040024, + 0.0013861599145457149, + 0.0012854456435889006, + 0.0022308758925646544, + 0.0018218866316601634, + 0.0011347418185323477, + 0.0021545032504945993, + 0.0016784091712906957, + 0.0021412603091448545, + 0.0038671810179948807, + 0.0029356263112276793, + 0.004073447547852993, + 0.0036985676269978285, + 0.0021729161962866783, + 0.0027426951564848423, + 0.0022289003245532513, + 0.0038034499157220125, + 0.004252094775438309, + 0.0028921852353960276, + 0.004453442990779877, + 0.004687745589762926, + 0.005300232674926519, + 0.004113843664526939, + 0.0034909010864794254, + 0.007234262768179178, + 0.005602213554084301, + 0.006711646448820829, + 0.00900548417121172, + 0.006528913509100676, + 0.00897268671542406, + 0.01200142316520214, + 0.015932178124785423, + 0.010008149780333042, + 0.014301070012152195, + 0.02405594103038311, + 0.08983512967824936, + 0.08505842089653015, + 0.027644282206892967, + 0.017869479954242706, + 0.020909331738948822, + 0.027921730652451515, + 0.04030245542526245, + 0.034567803144454956, + 0.17927737534046173, + 0.04869001358747482, + 0.17761969566345215, + 0.0017123740399256349, + 0.001759629463776946, + 0.0008765669772401452, + 0.0008780930656939745, + 0.0012041641166433692, + 0.0015345857245847583, + 0.001326192868873477, + 0.0015906200278550386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.018498403951525688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.01820182055234909 + ], + "opposite_logprob_tokens": [ + -4.01820182800293 + ], + "logprob_reward": -0.01820182055234909, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820182055234909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05028286203742027, + 0.02369716763496399, + 0.014071191661059856, + 0.01664765179157257, + 0.05790204182267189, + 0.05631665140390396, + 0.03750794008374214, + 0.3151198625564575, + 0.383079469203949, + 0.045375123620033264 + ], + "credit_map": [ + 0.0076991175301373005, + 0.002602000255137682, + 0.003935413900762796, + 0.003973213955760002, + 0.0050902049988508224, + 0.007965460419654846, + 0.004874732345342636, + 0.0141427181661129, + 0.0034473303239792585, + 0.0020619442220777273, + 0.003092353232204914, + 0.0031707328744232655, + 0.002279123058542609, + 0.002665863139554858, + 0.004285083618015051, + 0.0026947378646582365, + 0.0018358842935413122, + 0.0011761717032641172, + 0.002552782418206334, + 0.0014049329329282045, + 0.0013224835274741054, + 0.002308907685801387, + 0.0019020062172785401, + 0.001568024163134396, + 0.0017508264863863587, + 0.0015621442580595613, + 0.002562583889812231, + 0.002165274927392602, + 0.0019349263748154044, + 0.002475757384672761, + 0.00268371170386672, + 0.0015124279307201505, + 0.006643406115472317, + 0.006813975982367992, + 0.008730271831154823, + 0.007816982455551624, + 0.0063819242641329765, + 0.007037333678454161, + 0.009015301242470741, + 0.005462842993438244, + 0.005422649905085564, + 0.0029948772862553596, + 0.005523575469851494, + 0.00815681554377079, + 0.0044458345510065556, + 0.00806315615773201, + 0.009426173754036427, + 0.01228356920182705, + 0.002829392207786441, + 0.002269926480948925, + 0.0047883703373372555, + 0.0034890049137175083, + 0.005940128117799759, + 0.006056109443306923, + 0.005094417370855808, + 0.007040595170110464, + 0.013917534612119198, + 0.01310814544558525, + 0.009363547898828983, + 0.007438720669597387, + 0.02128962241113186, + 0.11461059749126434, + 0.10462667793035507, + 0.030765017494559288, + 0.012405957095324993, + 0.01403422188013792, + 0.021436519920825958, + 0.024883149191737175, + 0.022773368284106255, + 0.12849995493888855, + 0.03203696757555008, + 0.1270093023777008, + 0.004873848520219326, + 0.00449630431830883, + 0.006692191585898399, + 0.004745026584714651, + 0.006253290455788374, + 0.0050780256278812885, + 0.008100918494164944, + 0.0051355198957026005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.01820182055234909 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.04306260123848915 + ], + "opposite_logprob_tokens": [ + -3.168062686920166 + ], + "logprob_reward": -0.04306260123848915, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306260123848915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.097148597240448, + 0.03345018997788429, + 0.015680065378546715, + 0.029482584446668625, + 0.045080602169036865, + 0.09293972700834274, + 0.06088223308324814, + 0.1798378825187683, + 0.35790374875068665, + 0.08759436011314392 + ], + "credit_map": [ + 0.009373166598379612, + 0.005505407694727182, + 0.008399243466556072, + 0.009809048846364021, + 0.011458579450845718, + 0.019694510847330093, + 0.011922034434974194, + 0.02098660171031952, + 0.005053115077316761, + 0.0016783764585852623, + 0.004304502625018358, + 0.005807128269225359, + 0.0036666886880993843, + 0.003700748085975647, + 0.005450682248920202, + 0.0037889480590820312, + 0.0024683799128979445, + 0.0011313427239656448, + 0.0025203186087310314, + 0.0016584746772423387, + 0.001986492658033967, + 0.002027801936492324, + 0.0024319300428032875, + 0.0014553251676261425, + 0.003344521624967456, + 0.0031188151333481073, + 0.004189942963421345, + 0.004052425269037485, + 0.003389581572264433, + 0.004559182096272707, + 0.00444028852507472, + 0.0023878300562500954, + 0.005572587717324495, + 0.003860565135255456, + 0.004625148139894009, + 0.007031980901956558, + 0.004423112608492374, + 0.007981549948453903, + 0.006970797665417194, + 0.004614857956767082, + 0.005586985964328051, + 0.003651668783277273, + 0.005321100354194641, + 0.0068819220177829266, + 0.005155193619430065, + 0.026613015681505203, + 0.02835986018180847, + 0.01136997900903225, + 0.006298346444964409, + 0.004205895587801933, + 0.009760275483131409, + 0.007500484120100737, + 0.007996777072548866, + 0.00733517948538065, + 0.009293696843087673, + 0.00849157478660345, + 0.006934306584298611, + 0.01050436869263649, + 0.006966196466237307, + 0.009247361682355404, + 0.013082979246973991, + 0.05405939742922783, + 0.059403080493211746, + 0.019640188664197922, + 0.012319503352046013, + 0.012092451564967632, + 0.015216822735965252, + 0.015424404293298721, + 0.016401398926973343, + 0.13204994797706604, + 0.02518991194665432, + 0.12920932471752167, + 0.004668148700147867, + 0.004286269191652536, + 0.008649013005197048, + 0.008201293647289276, + 0.008786318823695183, + 0.01900121383368969, + 0.011934153735637665, + 0.02206794172525406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.04306260123848915 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.0046369642950594425 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.0046369642950594425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046369642950594425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08167527616024017, + 0.05475752428174019, + 0.02538912743330002, + 0.025406235828995705, + 0.04043295979499817, + 0.053742218762636185, + 0.053650714457035065, + 0.2371048480272293, + 0.3641175627708435, + 0.0637235939502716 + ], + "credit_map": [ + 0.013743260875344276, + 0.007170974742621183, + 0.01302399579435587, + 0.010573608800768852, + 0.010724806226789951, + 0.010387522168457508, + 0.007985669188201427, + 0.008065436966717243, + 0.006032455246895552, + 0.0048506492748856544, + 0.00785062275826931, + 0.007806594017893076, + 0.005706040654331446, + 0.007818340323865414, + 0.008382949978113174, + 0.006309868767857552, + 0.0029559137765318155, + 0.0024019586853682995, + 0.0035109680611640215, + 0.0033524532336741686, + 0.0026690938975661993, + 0.003145660273730755, + 0.004121819511055946, + 0.0032312595285475254, + 0.0026181270368397236, + 0.002790514612570405, + 0.004944115411490202, + 0.0032595782540738583, + 0.002419695956632495, + 0.003489553462713957, + 0.003803288796916604, + 0.002081360435113311, + 0.0038197075482457876, + 0.0025835451669991016, + 0.006195445079356432, + 0.006070861592888832, + 0.003918197471648455, + 0.007026598323136568, + 0.006715758703649044, + 0.004102844279259443, + 0.003938938491046429, + 0.0038073151372373104, + 0.005426143761724234, + 0.006171069573611021, + 0.004141680430620909, + 0.013614827767014503, + 0.011479943059384823, + 0.005162300541996956, + 0.0042447280138731, + 0.0035867346450686455, + 0.008212720975279808, + 0.005748922936618328, + 0.006687089800834656, + 0.010417929850518703, + 0.006437147501856089, + 0.008315433748066425, + 0.010278857313096523, + 0.012833203189074993, + 0.00824082363396883, + 0.011413519270718098, + 0.019547462463378906, + 0.07829165458679199, + 0.07332572340965271, + 0.023173602297902107, + 0.011753099039196968, + 0.016512705013155937, + 0.024143047630786896, + 0.03637012466788292, + 0.030711565166711807, + 0.10136029869318008, + 0.03969351574778557, + 0.10357318073511124, + 0.005278981290757656, + 0.0028009803500026464, + 0.003664276096969843, + 0.0038287572097033262, + 0.004039832390844822, + 0.019000867381691933, + 0.005018006544560194, + 0.020091889426112175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0046369642950594425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium close-up focused on a knight reclining against a massive cracked stone Buddha statue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the knight have long jet-black hair tied back with a cloth ribbon, stubble on his chin, and wear a frayed dark blue coarse cloth gown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knight's left hand resting on the hilt of an ancient dark-brown leather-wrapped longsword?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a cool clear moonlight pour through a gaping hole in the temple roof creating a shaft of light with dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the knight's eyes open suddenly and sharply from a closed state, staring directly at the camera with an intense 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": [ + -0.016069892793893814 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069892793893814, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069892793893814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04503367841243744, + 0.03468058258295059, + 0.02247415855526924, + 0.02560536563396454, + 0.07054111361503601, + 0.07619193196296692, + 0.044039536267519, + 0.2535766363143921, + 0.3313842713832855, + 0.09647279232740402 + ], + "credit_map": [ + 0.0069540985859930515, + 0.002272008452564478, + 0.004210127051919699, + 0.005351451691240072, + 0.005111663602292538, + 0.0063172513619065285, + 0.004430735949426889, + 0.010386339388787746, + 0.0044303471222519875, + 0.002412308007478714, + 0.003724271198734641, + 0.004552511498332024, + 0.006897075101733208, + 0.004996641539037228, + 0.004356643185019493, + 0.003310782602056861, + 0.0026831531431525946, + 0.0019203036790713668, + 0.005322306416928768, + 0.0025509630795568228, + 0.0023216726258397102, + 0.002251520287245512, + 0.0028264648281037807, + 0.002597773214802146, + 0.003204038366675377, + 0.0017477837391197681, + 0.0030615392606705427, + 0.003118944587185979, + 0.0029164673760533333, + 0.005349744111299515, + 0.004021767061203718, + 0.0021850785706192255, + 0.006284421309828758, + 0.004601108375936747, + 0.008508049882948399, + 0.01000183541327715, + 0.00742756575345993, + 0.013755488209426403, + 0.011573461815714836, + 0.008389171212911606, + 0.004201045259833336, + 0.003558254335075617, + 0.008469526655972004, + 0.008089450187981129, + 0.006003519054502249, + 0.019307570531964302, + 0.017043616622686386, + 0.0095189418643713, + 0.003333847736939788, + 0.002959162462502718, + 0.006857693661004305, + 0.005430647637695074, + 0.005911430809646845, + 0.00724055478349328, + 0.005786573980003595, + 0.006519623100757599, + 0.01232991460710764, + 0.016026847064495087, + 0.009241417050361633, + 0.013852007687091827, + 0.023743411526083946, + 0.08133277297019958, + 0.07253178209066391, + 0.024518471211194992, + 0.009941097348928452, + 0.01427678856998682, + 0.01959642767906189, + 0.029552273452281952, + 0.02764575369656086, + 0.09212496876716614, + 0.039796266704797745, + 0.09845063835382462, + 0.006022071931511164, + 0.0036728703416883945, + 0.0064517720602452755, + 0.005631845910102129, + 0.008214318193495274, + 0.027614712715148926, + 0.010572884231805801, + 0.028292309492826462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.016069892793893814 + } + ], + "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/0000110/group_4/index_tensor([1586, 1586])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_4/index_tensor([1586, 1586])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6527bcaef250252ee76cf6b1c9521701aeed1561 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_4/index_tensor([1586, 1586])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Adopting a soft, hand-drawn Miyazaki aesthetic, the camera executes smooth, low-angle sideways trajectories, gliding dreamily through lush, tall, swaying grass at golden hour. Each blade of grass, individually rendered with soft edges, glows vibrant greens and golds under the warm, syrupy sunlight, casting long, gentle shadows and imbuing the scene with a profound sense of nostalgia. A sudden, gentle breeze animates the frame, bending the grass into graceful waves and meticulously blowing swirling clouds of fluffy white goji berries into the air. Rendered as translucent, glowing spheres, these seeds float lazily and weightlessly across the frame, their slow, undulating journey creating a mesmerizing dance. As the camera continues its lateral movement, it reveals a forgotten metal kettle lying on its side, partially nestled in the grass, its once bright paint now faded and covered in patches of textured reddish-brown rust. A tiny ladybug crawls steadily and deliberately along the curve of the jar’s spout, its polished crimson shell and distinct black spot providing a sharp, vivid focal point as ethereal dragonfly seeds drift past, culminating in a quietly beautiful moment of natural wonder.", + "seed": 19552, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_4/index_1586_seed_19552_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": -0.026389128377195446, + "vlm_logprob": -0.026389128377195446, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.026389128377195446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004808054305613041 + ], + "opposite_logprob_tokens": [ + -7.254807949066162 + ], + "logprob_reward": -0.004808054305613041, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.004808054305613041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13583239912986755, + 0.09988149255514145, + 0.024647347629070282, + 0.010380130261182785, + 0.041174255311489105, + 0.012503470294177532, + 0.05050279572606087, + 0.03899485617876053, + 0.19961410760879517, + 0.3864690661430359 + ], + "credit_map": [ + 0.003719180589541793, + 0.05815843120217323, + 0.002648606663569808, + 0.0023004943504929543, + 0.008286654949188232, + 0.05598491430282593, + 0.0022641937248408794, + 0.0024699317291378975, + 0.0031861818861216307, + 0.043254248797893524, + 0.0014357821783050895, + 0.0013336790725588799, + 0.0418388769030571, + 0.0060095698572695255, + 0.0012971225660294294, + 0.0015260402578860521, + 0.0011238064616918564, + 0.006669970694929361, + 0.0011383595410734415, + 0.001432362594641745, + 0.0027512814849615097, + 0.007864530198276043, + 0.0019315107492730021, + 0.001735530560836196, + 0.0009524571360088885, + 0.0016657133819535375, + 0.0007540090009570122, + 0.0007574513438157737, + 0.0022688303142786026, + 0.0023511131294071674, + 0.0008289990946650505, + 0.0008015582570806146, + 0.002505678916350007, + 0.006978105753660202, + 0.002011419041082263, + 0.002517698798328638, + 0.012621439062058926, + 0.0072878520004451275, + 0.004907343070954084, + 0.002344724955037236, + 0.0009089928353205323, + 0.0015062288148328662, + 0.0010795175330713391, + 0.001425241120159626, + 0.0018379187677055597, + 0.0014436091296374798, + 0.0017154189990833402, + 0.002586544957011938, + 0.004752324428409338, + 0.008123071864247322, + 0.005687265191227198, + 0.00784993078559637, + 0.005340341944247484, + 0.008050213567912579, + 0.00578942010179162, + 0.004910231102257967, + 0.0018931603990495205, + 0.011904130689799786, + 0.0020100469700992107, + 0.0022354102693498135, + 0.012316332198679447, + 0.005482373759150505, + 0.0018503383034840226, + 0.0013030670816078782, + 0.017905931919813156, + 0.03260095417499542, + 0.02095290832221508, + 0.011831703595817089, + 0.0299997441470623, + 0.05849578604102135, + 0.011643501929938793, + 0.01618359237909317, + 0.0193457193672657, + 0.1631951630115509, + 0.012576279230415821, + 0.006357633974403143, + 0.02110166661441326, + 0.14236080646514893, + 0.004096637479960918, + 0.017435193061828613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.004808054305613041 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008191090892069042 + ], + "opposite_logprob_tokens": [ + -7.125819206237793 + ], + "logprob_reward": -0.0008191090892069042, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008191090892069042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0672394335269928, + 0.08884410560131073, + 0.011910908855497837, + 0.012404934503138065, + 0.016140436753630638, + 0.01591036282479763, + 0.036155518144369125, + 0.02316339686512947, + 0.05743500217795372, + 0.6707959175109863 + ], + "credit_map": [ + 0.00469661271199584, + 0.019752059131860733, + 0.0036598341539502144, + 0.0031307311728596687, + 0.007049424108117819, + 0.02206764928996563, + 0.0029968612361699343, + 0.0038862687069922686, + 0.002493142383173108, + 0.035873398184776306, + 0.001300761941820383, + 0.0011659548617899418, + 0.03902573883533478, + 0.006142801139503717, + 0.0013887303648516536, + 0.0014535754453390837, + 0.001426804345101118, + 0.002350372727960348, + 0.0013749717036262155, + 0.0010513176675885916, + 0.002635802375152707, + 0.001454987213946879, + 0.0008470452157780528, + 0.0007696080720052123, + 0.0012377232778817415, + 0.001758564030751586, + 0.0016651806654408574, + 0.0013521200744435191, + 0.0022007268853485584, + 0.0013689093757420778, + 0.0016144359251484275, + 0.0012072754325345159, + 0.0011513199424371123, + 0.002074589254334569, + 0.0009720380767248571, + 0.0013602470280602574, + 0.00402408093214035, + 0.0032964462880045176, + 0.0017919434467330575, + 0.0014697741717100143, + 0.0011115919332951307, + 0.0020690870005637407, + 0.001087347511202097, + 0.0017598727717995644, + 0.0036608115769922733, + 0.002166308695450425, + 0.0025023955386132, + 0.001552949077449739, + 0.00321116391569376, + 0.007105325814336538, + 0.0038660226855427027, + 0.0047239940613508224, + 0.006043564528226852, + 0.004806533921509981, + 0.003558962605893612, + 0.0028399534057825804, + 0.0012510239612311125, + 0.006687053479254246, + 0.0016846045618876815, + 0.002294920850545168, + 0.005624043755233288, + 0.002047835849225521, + 0.0017904387786984444, + 0.0017834785394370556, + 0.0050567747093737125, + 0.009065053425729275, + 0.006293370854109526, + 0.0036154701374471188, + 0.008092097006738186, + 0.01640898548066616, + 0.004026122856885195, + 0.0048771328292787075, + 0.03319569677114487, + 0.28348639607429504, + 0.02188056893646717, + 0.012089094147086143, + 0.037648506462574005, + 0.24560926854610443, + 0.007438673637807369, + 0.029447706416249275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0008191090892069042 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369152918457985 + ], + "opposite_logprob_tokens": [ + -3.408691644668579 + ], + "logprob_reward": -0.03369152918457985, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369152918457985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04021843150258064, + 0.025432873517274857, + 0.003758241655305028, + 0.006093150936067104, + 0.004965769127011299, + 0.007684358861297369, + 0.012373489327728748, + 0.010278872214257717, + 0.012425575405359268, + 0.8767692446708679 + ], + "credit_map": [ + 0.0015030851354822516, + 0.016491472721099854, + 0.0012279453221708536, + 0.0009336365037597716, + 0.0030865489970892668, + 0.014752180315554142, + 0.0011577581753954291, + 0.0010657970560714602, + 0.0009552636765874922, + 0.010388210415840149, + 0.0007281156722456217, + 0.0005365163669921458, + 0.009073212742805481, + 0.0019640610553324223, + 0.0006758845411241055, + 0.0011116061359643936, + 0.00041309656808152795, + 0.000723887060303241, + 0.000323642190778628, + 0.0003659750218503177, + 0.0007774030673317611, + 0.0004834581050090492, + 0.00031270034378394485, + 0.0003580785123631358, + 0.0006742468103766441, + 0.0008824669057503343, + 0.0008892451878637075, + 0.0005617782007902861, + 0.0009641333599574864, + 0.000900701095815748, + 0.0006031204829923809, + 0.0006174572627060115, + 0.0002859917585738003, + 0.0008072388591244817, + 0.0003183434600941837, + 0.0004516050103120506, + 0.0011551296338438988, + 0.0007213675417006016, + 0.0008927946910262108, + 0.00033329727011732757, + 0.0005410182639025152, + 0.0015548702795058489, + 0.0004838838358409703, + 0.0009766619186848402, + 0.001087637385353446, + 0.0010090464493259788, + 0.001182281062938273, + 0.0008489586762152612, + 0.001182583044283092, + 0.0020907847210764885, + 0.001116398605518043, + 0.0011193782556802034, + 0.003658109810203314, + 0.0014754871372133493, + 0.0009127574739977717, + 0.0008179882424883544, + 0.00039902818389236927, + 0.0035521509125828743, + 0.0005009634769521654, + 0.0007603462436236441, + 0.003380069974809885, + 0.0007596873911097646, + 0.0005500015686266124, + 0.0003766224253922701, + 0.0010789916850626469, + 0.001992245204746723, + 0.0013106776168569922, + 0.0008697052835486829, + 0.0018218971090391278, + 0.003598129376769066, + 0.000796301057562232, + 0.0009576267912052572, + 0.0435347706079483, + 0.37071338295936584, + 0.028321389108896255, + 0.014826700091362, + 0.0492624007165432, + 0.32136377692222595, + 0.010219004936516285, + 0.038527682423591614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.03369152918457985 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06663820147514343, + 0.08029042929410934, + 0.0391773097217083, + 0.008814120665192604, + 0.07521747052669525, + 0.04891018569469452, + 0.07159662991762161, + 0.038608670234680176, + 0.05046217143535614, + 0.52028489112854 + ], + "credit_map": [ + 0.002521543763577938, + 0.024181976914405823, + 0.002310710260644555, + 0.002226636279374361, + 0.004355594515800476, + 0.025037340819835663, + 0.0030681388452649117, + 0.002936257515102625, + 0.002121244091540575, + 0.03290063142776489, + 0.0011470599565654993, + 0.001032671774737537, + 0.035048868507146835, + 0.005450290162116289, + 0.0011619083816185594, + 0.001427755574695766, + 0.002162117976695299, + 0.014588217251002789, + 0.0021365168504416943, + 0.001941544353030622, + 0.0041631362400949, + 0.012180598452687263, + 0.0009461473673582077, + 0.001059028902091086, + 0.0010296072578057647, + 0.0013802212197333574, + 0.0011823067907243967, + 0.0009117116569541395, + 0.001593717490322888, + 0.0009532746043987572, + 0.00093693973030895, + 0.0008263420313596725, + 0.005818082019686699, + 0.013280542567372322, + 0.0051412214525043964, + 0.0064156511798501015, + 0.01974870078265667, + 0.013269971124827862, + 0.006913573946803808, + 0.00462972791865468, + 0.003934136126190424, + 0.00648034643381834, + 0.00399870565161109, + 0.007181432563811541, + 0.004031867254525423, + 0.005610410589724779, + 0.008185163140296936, + 0.00948812160640955, + 0.00603243475779891, + 0.009631799533963203, + 0.0057595837861299515, + 0.013893019407987595, + 0.01036486029624939, + 0.010069215670228004, + 0.009260306134819984, + 0.006585404276847839, + 0.0028815774712711573, + 0.007079474627971649, + 0.002668781904503703, + 0.00469372421503067, + 0.008344468660652637, + 0.0053751044906675816, + 0.005221042316406965, + 0.002344497013837099, + 0.004502568859606981, + 0.007590975612401962, + 0.004519474692642689, + 0.010466000065207481, + 0.004233944695442915, + 0.005321720149368048, + 0.009152860380709171, + 0.0046746255829930305, + 0.025677155703306198, + 0.21429629623889923, + 0.017858002334833145, + 0.01649412512779236, + 0.03137565031647682, + 0.18381139636039734, + 0.008665558882057667, + 0.02210668846964836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.003615511115640402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901143819093704 + ], + "opposite_logprob_tokens": [ + -2.4640114307403564 + ], + "logprob_reward": -0.08901143819093704, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901143819093704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06443790346384048, + 0.1408110409975052, + 0.026200497522950172, + 0.00855650007724762, + 0.07591631263494492, + 0.01757488027215004, + 0.01945807784795761, + 0.07040710002183914, + 0.178760826587677, + 0.3978768587112427 + ], + "credit_map": [ + 0.003977342043071985, + 0.022105742245912552, + 0.002138832351192832, + 0.0025642833206802607, + 0.005198509898036718, + 0.02260761894285679, + 0.0031750728376209736, + 0.002670502755790949, + 0.0038389936089515686, + 0.05956773832440376, + 0.001793819130398333, + 0.0016156123019754887, + 0.06134394183754921, + 0.00892470870167017, + 0.0018065054900944233, + 0.0019197110086679459, + 0.0013486734824255109, + 0.010373095981776714, + 0.0009765767026692629, + 0.0010500638745725155, + 0.0025624949485063553, + 0.008753242902457714, + 0.0005202927859500051, + 0.0006160552147775888, + 0.0011340293567627668, + 0.0012903421884402633, + 0.0011014475021511316, + 0.0008557250257581472, + 0.001431819866411388, + 0.0012642006622627378, + 0.0007285681203939021, + 0.0007503667729906738, + 0.005299726501107216, + 0.012389416806399822, + 0.00517163285985589, + 0.004762675613164902, + 0.02252979390323162, + 0.014367912895977497, + 0.0066703869961202145, + 0.0047247703187167645, + 0.0013855085708200932, + 0.0020723717752844095, + 0.001283509423956275, + 0.002303820103406906, + 0.002416989067569375, + 0.0020951898768544197, + 0.002629565540701151, + 0.0033879270777106285, + 0.0017441250383853912, + 0.0029547703452408314, + 0.0016886446392163634, + 0.0028909058310091496, + 0.0028104763478040695, + 0.003817716147750616, + 0.0018558388110250235, + 0.0016956005711108446, + 0.0034053290728479624, + 0.024547606706619263, + 0.0033188012894243, + 0.004148476291447878, + 0.023256300017237663, + 0.006198612041771412, + 0.00257752719335258, + 0.0029544434510171413, + 0.01582777500152588, + 0.0282460767775774, + 0.01924421824514866, + 0.011915454640984535, + 0.025699246674776077, + 0.051234107464551926, + 0.011219628155231476, + 0.015374318696558475, + 0.02007794938981533, + 0.16494689881801605, + 0.013400535099208355, + 0.010595646686851978, + 0.024216488003730774, + 0.14122538268566132, + 0.006841139402240515, + 0.01657284051179886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.08901143819093704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Adopting a soft, hand-drawn Miyazaki aesthetic, the camera executes smooth, low-angle sideways trajectories, gliding dreamily through lush, tall, swaying grass at golden hour. Each blade of grass, individually rendered with soft edges, glows vibrant greens and golds under the warm, syrupy sunlight, casting long, gentle shadows and imbuing the scene with a profound sense of nostalgia. A sudden, gentle breeze animates the frame, bending the grass into graceful waves and meticulously blowing swirling clouds of fluffy white goji berries into the air. Rendered as translucent, glowing spheres, these seeds float lazily and weightlessly across the frame, their slow, undulating journey creating a mesmerizing dance. As the camera continues its lateral movement, it reveals a forgotten metal kettle lying on its side, partially nestled in the grass, its once bright paint now faded and covered in patches of textured reddish-brown rust. A tiny ladybug crawls steadily and deliberately along the curve of the jar’s spout, its polished crimson shell and distinct black spot providing a sharp, vivid focal point as ethereal dragonfly seeds drift past, culminating in a quietly beautiful moment of natural wonder.", + "seed": 19553, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_4/index_1586_seed_19553_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": -0.03472579629160464, + "vlm_logprob": -0.03472579629160464, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03472579629160464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007742159068584442 + ], + "opposite_logprob_tokens": [ + -6.007741928100586 + ], + "logprob_reward": -0.007742159068584442, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007742159068584442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21636995673179626, + 0.1328781694173813, + 0.03465644270181656, + 0.18957091867923737, + 0.029049359261989594, + 0.10374011099338531, + 0.11441446840763092, + 0.07277138531208038, + 0.061589036136865616, + 0.04496011137962341 + ], + "credit_map": [ + 0.05755196884274483, + 0.08078570663928986, + 0.013766487129032612, + 0.013371246866881847, + 0.013390903361141682, + 0.015734273940324783, + 0.010190957225859165, + 0.011578411795198917, + 0.006572013255208731, + 0.02606615424156189, + 0.0069508678279817104, + 0.006734470836818218, + 0.04372773692011833, + 0.02639630250632763, + 0.009087882936000824, + 0.007342739496380091, + 0.00420342804864049, + 0.004739450290799141, + 0.0040710861794650555, + 0.004631460644304752, + 0.003988032694905996, + 0.005547931883484125, + 0.0037830600049346685, + 0.0036919964477419853, + 0.019855251535773277, + 0.02948363870382309, + 0.02068629488348961, + 0.019741276279091835, + 0.02505584806203842, + 0.04058697074651718, + 0.017732322216033936, + 0.016429316252470016, + 0.0024990644305944443, + 0.003183837281540036, + 0.0029152475763112307, + 0.005441597662866116, + 0.0020299996249377728, + 0.004163627978414297, + 0.005302066914737225, + 0.0035139170940965414, + 0.009254717268049717, + 0.010413638316094875, + 0.0107411565259099, + 0.01385430432856083, + 0.013298392295837402, + 0.017396792769432068, + 0.011124145239591599, + 0.01765696331858635, + 0.005183215253055096, + 0.006815158762037754, + 0.008023947477340698, + 0.006939368788152933, + 0.03475496545433998, + 0.039176203310489655, + 0.007011985406279564, + 0.006509631406515837, + 0.00800018198788166, + 0.007136488799005747, + 0.009224594570696354, + 0.006889614276587963, + 0.010890009813010693, + 0.011260158382356167, + 0.009292660281062126, + 0.01007767952978611, + 0.005780862178653479, + 0.009336541406810284, + 0.007481259759515524, + 0.010553165338933468, + 0.007833868265151978, + 0.008847769349813461, + 0.005858238786458969, + 0.005897332448512316, + 0.004689897410571575, + 0.0026885441038757563, + 0.00400724820792675, + 0.004348937422037125, + 0.0033424729481339455, + 0.013134554028511047, + 0.004173273220658302, + 0.008575188927352428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007742159068584442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002831618534401059 + ], + "opposite_logprob_tokens": [ + -5.87783145904541 + ], + "logprob_reward": -0.002831618534401059, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002831618534401059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2833233177661896, + 0.05950841307640076, + 0.06214341148734093, + 0.11999257653951645, + 0.03431547060608864, + 0.14354519546031952, + 0.16343210637569427, + 0.029973620548844337, + 0.03855316340923309, + 0.06521275639533997 + ], + "credit_map": [ + 0.08862397819757462, + 0.10541250556707382, + 0.016950709745287895, + 0.00938695203512907, + 0.016151798889040947, + 0.02080267108976841, + 0.012954339385032654, + 0.013040348887443542, + 0.004630983341485262, + 0.00874993298202753, + 0.004149200394749641, + 0.004810049664229155, + 0.01687268540263176, + 0.012304969131946564, + 0.004795110784471035, + 0.003195476485416293, + 0.00558279175311327, + 0.011725695803761482, + 0.007858442142605782, + 0.007050415966659784, + 0.009993831627070904, + 0.008554370142519474, + 0.0066116987727582455, + 0.004766163416206837, + 0.012654772028326988, + 0.020364221185445786, + 0.013140621595084667, + 0.013103070668876171, + 0.013693582266569138, + 0.023458024486899376, + 0.012829321436583996, + 0.010748970322310925, + 0.0032977531664073467, + 0.003164502792060375, + 0.004884212277829647, + 0.00415441207587719, + 0.005267298314720392, + 0.006248792167752981, + 0.00391193525865674, + 0.0033865650184452534, + 0.013533137738704681, + 0.012210020795464516, + 0.012837101705372334, + 0.013879372738301754, + 0.022404257208108902, + 0.029666857793927193, + 0.01173616573214531, + 0.027278289198875427, + 0.009242173284292221, + 0.007313880603760481, + 0.01058666780591011, + 0.008021608926355839, + 0.049516454339027405, + 0.057543013244867325, + 0.008782166987657547, + 0.01242613885551691, + 0.0030470089986920357, + 0.003973312210291624, + 0.0033707390539348125, + 0.0033078202977776527, + 0.004918772727251053, + 0.0051120794378221035, + 0.003793503390625119, + 0.0024503839667886496, + 0.004135688301175833, + 0.002911277813836932, + 0.0055628204718232155, + 0.005776780657470226, + 0.004239112138748169, + 0.007336599286645651, + 0.0042107305489480495, + 0.004380153492093086, + 0.008207914419472218, + 0.0057261185720562935, + 0.009660596027970314, + 0.012218296527862549, + 0.0038619027473032475, + 0.00984563585370779, + 0.006669553928077221, + 0.009022739715874195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.002831618534401059 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808869421482086 + ], + "opposite_logprob_tokens": [ + -3.288088798522949 + ], + "logprob_reward": -0.03808869421482086, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808869421482086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11000964045524597, + 0.08702405542135239, + 0.028931906446814537, + 0.031092509627342224, + 0.05430363118648529, + 0.3992817997932434, + 0.17994320392608643, + 0.028730709105730057, + 0.04937151446938515, + 0.03131110593676567 + ], + "credit_map": [ + 0.02049640566110611, + 0.020478561520576477, + 0.013530525378882885, + 0.010117405094206333, + 0.010030925273895264, + 0.011881633661687374, + 0.010554556734859943, + 0.012919630855321884, + 0.007828127592802048, + 0.00851333700120449, + 0.006490833126008511, + 0.006746090017259121, + 0.0238951425999403, + 0.020305553451180458, + 0.007103207055479288, + 0.006141765043139458, + 0.002753221895545721, + 0.0038701011799275875, + 0.0027885655872523785, + 0.0034297427628189325, + 0.004897579085081816, + 0.004390909802168608, + 0.0031983775552362204, + 0.003603406948968768, + 0.00351569801568985, + 0.004999575205147266, + 0.002702058292925358, + 0.0025046723894774914, + 0.006671071983873844, + 0.00537029467523098, + 0.002381497761234641, + 0.0029476387426257133, + 0.004638420417904854, + 0.004832910373806953, + 0.005831738002598286, + 0.010138588026165962, + 0.0062879049219191074, + 0.008945779874920845, + 0.008257887326180935, + 0.005370402708649635, + 0.035458266735076904, + 0.03223888576030731, + 0.030090730637311935, + 0.030718591064214706, + 0.04694012925028801, + 0.10667501389980316, + 0.02599913254380226, + 0.09116102010011673, + 0.0085506122559309, + 0.009384849108755589, + 0.011122469790279865, + 0.01232520304620266, + 0.05579069256782532, + 0.06115642562508583, + 0.011324499733746052, + 0.010288442485034466, + 0.0035717650316655636, + 0.00336279533803463, + 0.0026020980440080166, + 0.002354066353291273, + 0.004037952516227961, + 0.005250153131783009, + 0.0024488470517098904, + 0.005103033501654863, + 0.004659116733819246, + 0.0033914032392203808, + 0.006850137375295162, + 0.00814326573163271, + 0.0036948334891349077, + 0.007988489232957363, + 0.00491240993142128, + 0.009731859900057316, + 0.003398353233933449, + 0.003051977837458253, + 0.005439402535557747, + 0.006446264684200287, + 0.002955782925710082, + 0.0024834403302520514, + 0.0032832969445735216, + 0.004252586513757706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.03808869421482086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495113879442215 + ], + "opposite_logprob_tokens": [ + -2.9299511909484863 + ], + "logprob_reward": -0.05495113879442215, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495113879442215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03429875895380974, + 0.06264875829219818, + 0.006509443279355764, + 0.09164509922266006, + 0.04753004387021065, + 0.08347594738006592, + 0.2888600528240204, + 0.09488276392221451, + 0.23663894832134247, + 0.053510189056396484 + ], + "credit_map": [ + 0.007385934703052044, + 0.008859983645379543, + 0.002564922906458378, + 0.002593273762613535, + 0.0033875652588903904, + 0.0037337446119636297, + 0.002805393887683749, + 0.0029679390136152506, + 0.0038620266132056713, + 0.00947671476751566, + 0.0030966659542173147, + 0.0038302424363791943, + 0.02083887718617916, + 0.013790986500680447, + 0.004089261405169964, + 0.003663986688479781, + 0.0009319581440649927, + 0.0009484935435466468, + 0.0007283692248165607, + 0.0006892651435919106, + 0.0010477748000994325, + 0.0009045038605108857, + 0.0006678677746094763, + 0.0005912110791541636, + 0.009143211878836155, + 0.014531473629176617, + 0.011256693862378597, + 0.009043153375387192, + 0.011978392489254475, + 0.01991797238588333, + 0.008031726814806461, + 0.007742467802017927, + 0.003659779205918312, + 0.004402989521622658, + 0.004544014576822519, + 0.003977905958890915, + 0.006046158727258444, + 0.011037707328796387, + 0.004478190094232559, + 0.009383298456668854, + 0.00791662186384201, + 0.007207500282675028, + 0.006728438660502434, + 0.006821165792644024, + 0.009499825537204742, + 0.020199015736579895, + 0.00708008324727416, + 0.018023289740085602, + 0.018517209216952324, + 0.026552535593509674, + 0.023207375779747963, + 0.031310420483350754, + 0.07436720281839371, + 0.06556966155767441, + 0.03066842071712017, + 0.01866724342107773, + 0.008927999995648861, + 0.01246707234531641, + 0.011902283877134323, + 0.010956878773868084, + 0.013253534212708473, + 0.013991044834256172, + 0.011148017831146717, + 0.01223592646420002, + 0.015426905825734138, + 0.01559491641819477, + 0.017847763374447823, + 0.024472560733556747, + 0.01835528016090393, + 0.07579308748245239, + 0.01409724447876215, + 0.05505118891596794, + 0.005376407876610756, + 0.004933314863592386, + 0.008982731029391289, + 0.01071219053119421, + 0.004029867239296436, + 0.006851696874946356, + 0.0056470585986971855, + 0.006976922508329153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.05495113879442215 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001537084579468 + ], + "opposite_logprob_tokens": [ + -2.6950154304504395 + ], + "logprob_reward": -0.07001537084579468, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001537084579468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1127728745341301, + 0.028851842507719994, + 0.017335278913378716, + 0.05496210977435112, + 0.024158645421266556, + 0.2125832885503769, + 0.23103709518909454, + 0.09052211791276932, + 0.1091284528374672, + 0.11864831298589706 + ], + "credit_map": [ + 0.026565883308649063, + 0.03647990897297859, + 0.007173186168074608, + 0.006170029751956463, + 0.012667101807892323, + 0.010381855070590973, + 0.004686705768108368, + 0.008648204617202282, + 0.002420804463326931, + 0.0037626635748893023, + 0.00192833854816854, + 0.002112572779878974, + 0.007103376090526581, + 0.005788679234683514, + 0.002825846429914236, + 0.0029095602221786976, + 0.0018511021044105291, + 0.0015810454497113824, + 0.0015238258056342602, + 0.0015887303743511438, + 0.0034898328594863415, + 0.0042534153908491135, + 0.0017417859053239226, + 0.0013055411400273442, + 0.005472797434777021, + 0.008811774663627148, + 0.00649634562432766, + 0.00662170210853219, + 0.006461242213845253, + 0.01035449281334877, + 0.005510880146175623, + 0.005232872441411018, + 0.0017603154992684722, + 0.0022564867977052927, + 0.0024216121528297663, + 0.004383910447359085, + 0.0025234457571059465, + 0.0033740815706551075, + 0.0037958191242069006, + 0.0036429734900593758, + 0.019090089946985245, + 0.01738322339951992, + 0.016812968999147415, + 0.016702769324183464, + 0.024799151346087456, + 0.05626160651445389, + 0.013674029149115086, + 0.04785944148898125, + 0.012709582224488258, + 0.01242209505289793, + 0.015970613807439804, + 0.02141941525042057, + 0.055182792246341705, + 0.07314164191484451, + 0.02043497934937477, + 0.019755976274609566, + 0.008480556309223175, + 0.011421043425798416, + 0.005003728438168764, + 0.005988586228340864, + 0.023890048265457153, + 0.0193636491894722, + 0.008961953222751617, + 0.007412556558847427, + 0.0073704966343939304, + 0.007133364211767912, + 0.012044207192957401, + 0.01599876582622528, + 0.008715465664863586, + 0.029523374512791634, + 0.00914300512522459, + 0.01919977180659771, + 0.01056555937975645, + 0.009156319312751293, + 0.016461476683616638, + 0.021740566939115524, + 0.010492793284356594, + 0.0212284903973341, + 0.012224757112562656, + 0.01677834615111351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.07001537084579468 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Adopting a soft, hand-drawn Miyazaki aesthetic, the camera executes smooth, low-angle sideways trajectories, gliding dreamily through lush, tall, swaying grass at golden hour. Each blade of grass, individually rendered with soft edges, glows vibrant greens and golds under the warm, syrupy sunlight, casting long, gentle shadows and imbuing the scene with a profound sense of nostalgia. A sudden, gentle breeze animates the frame, bending the grass into graceful waves and meticulously blowing swirling clouds of fluffy white goji berries into the air. Rendered as translucent, glowing spheres, these seeds float lazily and weightlessly across the frame, their slow, undulating journey creating a mesmerizing dance. As the camera continues its lateral movement, it reveals a forgotten metal kettle lying on its side, partially nestled in the grass, its once bright paint now faded and covered in patches of textured reddish-brown rust. A tiny ladybug crawls steadily and deliberately along the curve of the jar’s spout, its polished crimson shell and distinct black spot providing a sharp, vivid focal point as ethereal dragonfly seeds drift past, culminating in a quietly beautiful moment of natural wonder.", + "seed": 19554, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_4/index_1586_seed_19554_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": -0.013630715338513255, + "vlm_logprob": -0.013630715338513255, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.013630715338513255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004906400106847286 + ], + "opposite_logprob_tokens": [ + -6.379906177520752 + ], + "logprob_reward": -0.004906400106847286, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.004906400106847286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20835721492767334, + 0.0567319393157959, + 0.09290260076522827, + 0.04039064049720764, + 0.06843598932027817, + 0.05766623094677925, + 0.26985836029052734, + 0.0311350729316473, + 0.09222717583179474, + 0.08229470252990723 + ], + "credit_map": [ + 0.016636673361063004, + 0.03333160653710365, + 0.014769446104764938, + 0.01129158679395914, + 0.03669019043445587, + 0.05760808289051056, + 0.022605646401643753, + 0.015424001961946487, + 0.005544370971620083, + 0.01084082666784525, + 0.0061999736353755, + 0.005803667940199375, + 0.007078675087541342, + 0.012162790633738041, + 0.004990202374756336, + 0.00411143759265542, + 0.007723629474639893, + 0.006215049419552088, + 0.004365584347397089, + 0.0038807762321084738, + 0.034036148339509964, + 0.029898185282945633, + 0.0035529891029000282, + 0.003230249509215355, + 0.0037150310818105936, + 0.0038701053708791733, + 0.0030293420422822237, + 0.005025162827223539, + 0.008751869201660156, + 0.007839524187147617, + 0.004247921518981457, + 0.003911686595529318, + 0.006290956400334835, + 0.010087591595947742, + 0.0038588831666857004, + 0.007488679140806198, + 0.01287008635699749, + 0.012355071492493153, + 0.006820628419518471, + 0.008664099499583244, + 0.005239126738160849, + 0.010651174001395702, + 0.006231322884559631, + 0.008725117892026901, + 0.00895520206540823, + 0.00796520709991455, + 0.0050628711469471455, + 0.00483621284365654, + 0.02112533338367939, + 0.037998054176568985, + 0.032339341938495636, + 0.034181803464889526, + 0.03262951970100403, + 0.03375619277358055, + 0.0416475348174572, + 0.03618061542510986, + 0.0030993663240224123, + 0.004092798568308353, + 0.0038826665841042995, + 0.0044683730229735374, + 0.005858982913196087, + 0.004120918456465006, + 0.0034639479126781225, + 0.0021480200812220573, + 0.006468535866588354, + 0.010444375686347485, + 0.01682461053133011, + 0.004022177308797836, + 0.03478333353996277, + 0.010714461095631123, + 0.003953111357986927, + 0.005016574170440435, + 0.008013749495148659, + 0.025578483939170837, + 0.00763928284868598, + 0.007531959563493729, + 0.009061540476977825, + 0.009541843086481094, + 0.006944884080439806, + 0.007982964627444744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.004906400106847286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009267802815884352 + ], + "opposite_logprob_tokens": [ + -7.000926971435547 + ], + "logprob_reward": -0.0009267802815884352, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009267802815884352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39208945631980896, + 0.03372366726398468, + 0.12790554761886597, + 0.09465324133634567, + 0.03373655304312706, + 0.04484491050243378, + 0.06444492191076279, + 0.03344395384192467, + 0.1293799877166748, + 0.04577776417136192 + ], + "credit_map": [ + 0.029150914400815964, + 0.060126472264528275, + 0.032298412173986435, + 0.026451773941516876, + 0.06537866592407227, + 0.10234468430280685, + 0.04243076965212822, + 0.033907756209373474, + 0.003961639944463968, + 0.004142398945987225, + 0.006327339448034763, + 0.0038653267547488213, + 0.0040824138559401035, + 0.004258741158992052, + 0.003984253387898207, + 0.003101552836596966, + 0.03158121928572655, + 0.011584480293095112, + 0.005601859651505947, + 0.006852075923234224, + 0.01686001941561699, + 0.04387836903333664, + 0.0034731621854007244, + 0.008074367418885231, + 0.0058325049467384815, + 0.007693993393331766, + 0.00450526550412178, + 0.007520316634327173, + 0.027220703661441803, + 0.030055511742830276, + 0.006191215477883816, + 0.0056337276473641396, + 0.004022012464702129, + 0.006529647391289473, + 0.002393522998318076, + 0.0036690994165837765, + 0.0037754883524030447, + 0.004924225155264139, + 0.003845047205686569, + 0.004577509127557278, + 0.004066561348736286, + 0.008864847011864185, + 0.004811795428395271, + 0.005891524720937014, + 0.00815477967262268, + 0.00663673086091876, + 0.0035840037744492292, + 0.002834668615832925, + 0.005599529016762972, + 0.008336082100868225, + 0.006372669246047735, + 0.007798846811056137, + 0.009261632338166237, + 0.010007960721850395, + 0.008845919743180275, + 0.008222278207540512, + 0.0032856201287359, + 0.005063456017524004, + 0.0038158332463353872, + 0.003440911415964365, + 0.007184031885117292, + 0.00370456138625741, + 0.0034669453743845224, + 0.0034825955517590046, + 0.007865468971431255, + 0.017537925392389297, + 0.016568075865507126, + 0.0054407245479524136, + 0.054379772394895554, + 0.016098398715257645, + 0.005191750358790159, + 0.006297870073467493, + 0.004205054137855768, + 0.012445975095033646, + 0.006798626855015755, + 0.0029222397133708, + 0.004178992938250303, + 0.007996848784387112, + 0.0038795878645032644, + 0.0033504378516227007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0009267802815884352 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532881461083889 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532881461083889, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532881461083889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21866323053836823, + 0.047321297228336334, + 0.06557603925466537, + 0.08255452662706375, + 0.046975575387477875, + 0.04049300402402878, + 0.3054780960083008, + 0.05282993242144585, + 0.049743786454200745, + 0.09036453068256378 + ], + "credit_map": [ + 0.013575959950685501, + 0.03151680901646614, + 0.018719177693128586, + 0.022468388080596924, + 0.02909667044878006, + 0.05910969898104668, + 0.021299652755260468, + 0.022876854985952377, + 0.004251709673553705, + 0.00865241326391697, + 0.005354300606995821, + 0.004118144046515226, + 0.005656025372445583, + 0.010965926572680473, + 0.003922856878489256, + 0.004399924073368311, + 0.011107238009572029, + 0.005194773897528648, + 0.0027362436521798372, + 0.002933819778263569, + 0.014230499044060707, + 0.02303675003349781, + 0.002187072765082121, + 0.004149637185037136, + 0.0056502860970795155, + 0.008008881472051144, + 0.0042632040567696095, + 0.006322093307971954, + 0.023562103509902954, + 0.02242933213710785, + 0.006707626394927502, + 0.005611003376543522, + 0.004799604415893555, + 0.005313549656420946, + 0.004843154922127724, + 0.007898764684796333, + 0.00628312211483717, + 0.007233652286231518, + 0.005938949063420296, + 0.004664778709411621, + 0.0034340505953878164, + 0.005632149055600166, + 0.004581157583743334, + 0.006472633220255375, + 0.006973508279770613, + 0.007709999568760395, + 0.0032364025246351957, + 0.002453102497383952, + 0.021672047674655914, + 0.04008949548006058, + 0.038735195994377136, + 0.03875277563929558, + 0.03654185310006142, + 0.04319196566939354, + 0.044878143817186356, + 0.041616614907979965, + 0.005142423789948225, + 0.007733110338449478, + 0.0074023958295583725, + 0.005209104623645544, + 0.013263552449643612, + 0.007366667501628399, + 0.004163847770541906, + 0.0025488315150141716, + 0.004857705906033516, + 0.005708308424800634, + 0.008132190443575382, + 0.0045971954241395, + 0.013692605309188366, + 0.005268608219921589, + 0.003852569730952382, + 0.0036346029955893755, + 0.007520320825278759, + 0.026611676439642906, + 0.008995219133794308, + 0.008818056434392929, + 0.011263585649430752, + 0.012541815638542175, + 0.006646359339356422, + 0.007967500016093254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.012532881461083889 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733820773661137 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006733820773661137, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733820773661137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10366511344909668, + 0.01740516908466816, + 0.033375926315784454, + 0.04779816046357155, + 0.03522014990448952, + 0.04535256698727608, + 0.3010353446006775, + 0.05167141929268837, + 0.27986466884613037, + 0.08461149036884308 + ], + "credit_map": [ + 0.007673176936805248, + 0.018647288903594017, + 0.010211221873760223, + 0.010765929706394672, + 0.015133802779018879, + 0.02047782763838768, + 0.00895808544009924, + 0.011797776445746422, + 0.0017200092552229762, + 0.00362895498983562, + 0.0021335489582270384, + 0.0015249170828610659, + 0.002727495040744543, + 0.0024618986062705517, + 0.0017574807861819863, + 0.0014508648309856653, + 0.004556215833872557, + 0.00460809376090765, + 0.002005490707233548, + 0.0028895377181470394, + 0.007572000380605459, + 0.007499803323298693, + 0.0021454847883433104, + 0.0020992993377149105, + 0.0030770248267799616, + 0.0058010597713291645, + 0.002540954854339361, + 0.004366899374872446, + 0.012932788580656052, + 0.011339004151523113, + 0.004388654604554176, + 0.003351778257638216, + 0.003784626955166459, + 0.004610174335539341, + 0.0030177230946719646, + 0.003736276878044009, + 0.004734731279313564, + 0.0044882758520543575, + 0.004931186325848103, + 0.005917156580835581, + 0.0040181465446949005, + 0.012655370868742466, + 0.0052510034292936325, + 0.004456833936274052, + 0.00814882293343544, + 0.006002368871122599, + 0.0026448583230376244, + 0.0021751616150140762, + 0.021242935210466385, + 0.038749292492866516, + 0.036960359662771225, + 0.039081357419490814, + 0.03551711514592171, + 0.04420359432697296, + 0.04407770186662674, + 0.04120299592614174, + 0.0034041376784443855, + 0.006740535609424114, + 0.0052168299444019794, + 0.0071641309186816216, + 0.013610202819108963, + 0.005073654931038618, + 0.006222617346793413, + 0.004239310044795275, + 0.019211897626519203, + 0.0382261648774147, + 0.03051372617483139, + 0.018459271639585495, + 0.11305136978626251, + 0.03408613055944443, + 0.012238177470862865, + 0.01407792791724205, + 0.007613359484821558, + 0.014961143024265766, + 0.012763815931975842, + 0.010464702732861042, + 0.0094413748010993, + 0.008811554871499538, + 0.010245071724057198, + 0.010310471057891846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.006733820773661137 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305369406938553 + ], + "opposite_logprob_tokens": [ + -3.16805362701416 + ], + "logprob_reward": -0.04305369406938553, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305369406938553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03632935509085655, + 0.029223628342151642, + 0.0361487939953804, + 0.013742013834416866, + 0.03162573277950287, + 0.034502483904361725, + 0.6311986446380615, + 0.06426236033439636, + 0.08177027106285095, + 0.04119670391082764 + ], + "credit_map": [ + 0.004402184393256903, + 0.005059889517724514, + 0.004217019770294428, + 0.0026003497187048197, + 0.0048258197493851185, + 0.007510430179536343, + 0.0036132147070020437, + 0.004100445192307234, + 0.003035297617316246, + 0.006048348266631365, + 0.0017137827817350626, + 0.0022027899976819754, + 0.0034451643005013466, + 0.007396789733320475, + 0.0029370859265327454, + 0.002444368787109852, + 0.002188372425734997, + 0.0029643739107996225, + 0.0016463359352201223, + 0.0015823880676180124, + 0.012885985895991325, + 0.011341217905282974, + 0.002135321730747819, + 0.001404798123985529, + 0.0010535792680457234, + 0.0016192678594961762, + 0.0011369952699169517, + 0.0016870874678716063, + 0.002919209422543645, + 0.003014259273186326, + 0.0011890887981280684, + 0.0011225261259824038, + 0.0020865367259830236, + 0.004456014838069677, + 0.0019244819413870573, + 0.0034204625990241766, + 0.00704858498647809, + 0.00628933310508728, + 0.0036673571448773146, + 0.0027329602744430304, + 0.003019830910488963, + 0.004438674543052912, + 0.0029180534183979034, + 0.0055186692625284195, + 0.008441837504506111, + 0.00471050338819623, + 0.003424056340008974, + 0.002030857838690281, + 0.044898103922605515, + 0.07544203102588654, + 0.06740286946296692, + 0.0815255418419838, + 0.07739794254302979, + 0.11400678008794785, + 0.086466945707798, + 0.08405846357345581, + 0.007196791470050812, + 0.00785689614713192, + 0.008096695877611637, + 0.009829366579651833, + 0.013545049354434013, + 0.009193304926156998, + 0.0049754902720451355, + 0.003568765940144658, + 0.0053196256048977375, + 0.011766236275434494, + 0.007519313134253025, + 0.012591956183314323, + 0.01873704046010971, + 0.009417482651770115, + 0.007330587133765221, + 0.00908802542835474, + 0.0033974433317780495, + 0.006091002840548754, + 0.005907506216317415, + 0.004674247931689024, + 0.0049068508669734, + 0.00561873335391283, + 0.006486549973487854, + 0.0041143703274428844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.04305369406938553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Adopting a soft, hand-drawn Miyazaki aesthetic, the camera executes smooth, low-angle sideways trajectories, gliding dreamily through lush, tall, swaying grass at golden hour. Each blade of grass, individually rendered with soft edges, glows vibrant greens and golds under the warm, syrupy sunlight, casting long, gentle shadows and imbuing the scene with a profound sense of nostalgia. A sudden, gentle breeze animates the frame, bending the grass into graceful waves and meticulously blowing swirling clouds of fluffy white goji berries into the air. Rendered as translucent, glowing spheres, these seeds float lazily and weightlessly across the frame, their slow, undulating journey creating a mesmerizing dance. As the camera continues its lateral movement, it reveals a forgotten metal kettle lying on its side, partially nestled in the grass, its once bright paint now faded and covered in patches of textured reddish-brown rust. A tiny ladybug crawls steadily and deliberately along the curve of the jar’s spout, its polished crimson shell and distinct black spot providing a sharp, vivid focal point as ethereal dragonfly seeds drift past, culminating in a quietly beautiful moment of natural wonder.", + "seed": 19555, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_4/index_1586_seed_19555_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": -0.017724284250289202, + "vlm_logprob": -0.017724284250289202, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.017724284250289202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011240961961448193 + ], + "opposite_logprob_tokens": [ + -4.8862409591674805 + ], + "logprob_reward": -0.011240961961448193, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011240961961448193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07262973487377167, + 0.09054026007652283, + 0.1919620782136917, + 0.31945887207984924, + 0.03147353231906891, + 0.033597223460674286, + 0.20563170313835144, + 0.013407590799033642, + 0.0332147441804409, + 0.00808422826230526 + ], + "credit_map": [ + 0.00637167040258646, + 0.007711309939622879, + 0.023254169151186943, + 0.007734761107712984, + 0.008632197044789791, + 0.007921579293906689, + 0.005550839938223362, + 0.005453211721032858, + 0.0069605037569999695, + 0.009902726858854294, + 0.015583235770463943, + 0.010942259803414345, + 0.008519756607711315, + 0.011545845307409763, + 0.011934265494346619, + 0.015151668339967728, + 0.04747609421610832, + 0.04042567312717438, + 0.016359737142920494, + 0.011200468987226486, + 0.020962173119187355, + 0.024207254871726036, + 0.014099875465035439, + 0.01723080314695835, + 0.02353092096745968, + 0.12013950943946838, + 0.009347078390419483, + 0.01092416513711214, + 0.014750641770660877, + 0.11457231640815735, + 0.007944821380078793, + 0.01824943721294403, + 0.003655077191069722, + 0.003845124039798975, + 0.004194777458906174, + 0.003485176945105195, + 0.003468624781817198, + 0.004512782674282789, + 0.003955423831939697, + 0.00435654679313302, + 0.004279258195310831, + 0.0031013323459774256, + 0.0038846901152282953, + 0.0024808146990835667, + 0.0033670156262815, + 0.00793877337127924, + 0.003860417054966092, + 0.00468492554500699, + 0.020345289260149002, + 0.0037730466574430466, + 0.006692945025861263, + 0.00500249769538641, + 0.007702842820435762, + 0.08601539582014084, + 0.012654755264520645, + 0.06344493478536606, + 0.002110840752720833, + 0.0010259306291118264, + 0.0012673307210206985, + 0.001846783678047359, + 0.0010793189285323024, + 0.0021051724907010794, + 0.002321042353287339, + 0.0016511721769347787, + 0.0049859085120260715, + 0.0018516193376854062, + 0.0038576326332986355, + 0.0030110429506748915, + 0.0022758154664188623, + 0.006841221358627081, + 0.003399101784452796, + 0.006992405280470848, + 0.0009052241803146899, + 0.0005656385328620672, + 0.0005349185667000711, + 0.0016127737471833825, + 0.0006192442961037159, + 0.00110229745041579, + 0.0017449839506298304, + 0.0009991488186642528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.011240961961448193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741635967046022 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741635967046022, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741635967046022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07468793541193008, + 0.06796765327453613, + 0.15205034613609314, + 0.3237588107585907, + 0.012644575908780098, + 0.018474090844392776, + 0.30980029702186584, + 0.01574508659541607, + 0.020885735750198364, + 0.003985450603067875 + ], + "credit_map": [ + 0.008562769740819931, + 0.009124336764216423, + 0.021285833790898323, + 0.008640863932669163, + 0.0073562380857765675, + 0.00946172047406435, + 0.005043352488428354, + 0.005212816875427961, + 0.005143355578184128, + 0.006621300242841244, + 0.01083566527813673, + 0.007579798810184002, + 0.009728874079883099, + 0.01096546370536089, + 0.007326209452003241, + 0.009766981936991215, + 0.04092422500252724, + 0.038203492760658264, + 0.012812352739274502, + 0.009427045471966267, + 0.017367679625749588, + 0.014115352183580399, + 0.00859619677066803, + 0.010603981092572212, + 0.02468889206647873, + 0.12433493137359619, + 0.008659803308546543, + 0.00945279374718666, + 0.01467451173812151, + 0.11637996137142181, + 0.007552269846200943, + 0.018015654757618904, + 0.0010016808519139886, + 0.0014649758813902736, + 0.002182236174121499, + 0.0014195325784385204, + 0.001557924784719944, + 0.001869985251687467, + 0.0016403222689405084, + 0.001507918001152575, + 0.0023361998610198498, + 0.0017547509633004665, + 0.0019575755577534437, + 0.0015434755478054285, + 0.001508893445134163, + 0.003842916339635849, + 0.002242600079625845, + 0.0032876792829483747, + 0.03288871794939041, + 0.005062729120254517, + 0.00927333440631628, + 0.005601937882602215, + 0.01150503195822239, + 0.13462847471237183, + 0.01689419336616993, + 0.09394588321447372, + 0.0029591908678412437, + 0.0008926396258175373, + 0.0012672250159084797, + 0.002024357905611396, + 0.001029453007504344, + 0.003088923403993249, + 0.0021606883965432644, + 0.002322607208043337, + 0.0029606043826788664, + 0.0010788800427690148, + 0.00272219511680305, + 0.001772568910382688, + 0.0013948832638561726, + 0.004283667542040348, + 0.001909067970700562, + 0.004763868171721697, + 0.000637647055555135, + 0.00030606245854869485, + 0.00030799544765613973, + 0.0004697145486716181, + 0.0004389365785755217, + 0.0006161649362184107, + 0.0006496027926914394, + 0.00055932707618922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.006741635967046022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011065899394452572 + ], + "opposite_logprob_tokens": [ + -4.51106595993042 + ], + "logprob_reward": -0.011065899394452572, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011065899394452572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06835679709911346, + 0.06610975414514542, + 0.1320873200893402, + 0.35758522152900696, + 0.017054138705134392, + 0.058535248041152954, + 0.2512146234512329, + 0.021963920444250107, + 0.017139434814453125, + 0.009953545406460762 + ], + "credit_map": [ + 0.008394621312618256, + 0.006896594539284706, + 0.016884390264749527, + 0.0074769300408661366, + 0.007026865612715483, + 0.011560924351215363, + 0.00439975643530488, + 0.005716710351407528, + 0.004125744570046663, + 0.0059477840550243855, + 0.0093096187338233, + 0.007627362385392189, + 0.010431083850562572, + 0.009821808896958828, + 0.008062412030994892, + 0.010783934965729713, + 0.035245005041360855, + 0.03467756509780884, + 0.012397789396345615, + 0.008295705541968346, + 0.01447597611695528, + 0.010566437616944313, + 0.006819125730544329, + 0.009609704837203026, + 0.02379292994737625, + 0.12955351173877716, + 0.009614049457013607, + 0.011076526716351509, + 0.017234699800610542, + 0.13470743596553802, + 0.008782828226685524, + 0.02282322570681572, + 0.001946972799487412, + 0.0021495737601071596, + 0.0019281923305243254, + 0.001812861650250852, + 0.001980777131393552, + 0.0025102721992880106, + 0.0024620816111564636, + 0.0022634060587733984, + 0.006970854010432959, + 0.0036345592234283686, + 0.004801791161298752, + 0.00435993168503046, + 0.003947730641812086, + 0.016453159973025322, + 0.005766191519796848, + 0.012601029127836227, + 0.03156813234090805, + 0.0047812084667384624, + 0.008388541638851166, + 0.005440619774162769, + 0.01063561998307705, + 0.11111727356910706, + 0.013067505322396755, + 0.06621573120355606, + 0.0035958020016551018, + 0.0012996990699321032, + 0.0015414607478305697, + 0.0032483807299286127, + 0.0017802256625145674, + 0.0030767107382416725, + 0.004424507264047861, + 0.0029971320182085037, + 0.0026995434891432524, + 0.0011883156839758158, + 0.0020851788576692343, + 0.002252373844385147, + 0.001324953860603273, + 0.002648785477504134, + 0.0023494837805628777, + 0.002590800169855356, + 0.0021535204723477364, + 0.0006587270181626081, + 0.0007225472363643348, + 0.0011654376285150647, + 0.0009257413912564516, + 0.001374043058604002, + 0.0016646375879645348, + 0.0012888896744698286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.011065899394452572 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787329956889153 + ], + "opposite_logprob_tokens": [ + -3.529787302017212 + ], + "logprob_reward": -0.029787329956889153, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787329956889153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038063209503889084, + 0.018124982714653015, + 0.04709089919924736, + 0.0824362263083458, + 0.011552390642464161, + 0.04481487348675728, + 0.6862824559211731, + 0.03530343621969223, + 0.010572369210422039, + 0.025759145617485046 + ], + "credit_map": [ + 0.0037120014894753695, + 0.0032740922179073095, + 0.00958833284676075, + 0.004841311369091272, + 0.003965229261666536, + 0.006184415426105261, + 0.0030093661043792963, + 0.0034884619526565075, + 0.00167980941478163, + 0.0018033150117844343, + 0.004468996077775955, + 0.0023410243447870016, + 0.0021095897536724806, + 0.0019321762956678867, + 0.0018600445473566651, + 0.0019300268031656742, + 0.01119306031614542, + 0.009428695775568485, + 0.003715297905728221, + 0.002964325249195099, + 0.005417488981038332, + 0.005422255955636501, + 0.004750798922032118, + 0.004198977258056402, + 0.006349264178425074, + 0.029593782499432564, + 0.004201974254101515, + 0.003814185503870249, + 0.004454811569303274, + 0.026368502527475357, + 0.003532747970893979, + 0.004120956175029278, + 0.0014605793403461576, + 0.0013075316091999412, + 0.0013755870750173926, + 0.0012541335308924317, + 0.0012458015698939562, + 0.0015245323302224278, + 0.0015702940290793777, + 0.0018139317398890853, + 0.005316136404871941, + 0.0026048726867884398, + 0.0037783535663038492, + 0.0034096285235136747, + 0.002762517658993602, + 0.011969013139605522, + 0.004963488783687353, + 0.010010860860347748, + 0.07252570986747742, + 0.011576106771826744, + 0.020061012357473373, + 0.012572572566568851, + 0.025325503200292587, + 0.2982904314994812, + 0.03762531653046608, + 0.20830580592155457, + 0.003412974067032337, + 0.002837021602317691, + 0.0030039683915674686, + 0.008728663437068462, + 0.002745003905147314, + 0.004606582690030336, + 0.0066756997257471085, + 0.0032935235649347305, + 0.0014324432704597712, + 0.0005902094999328256, + 0.001109523931518197, + 0.001026882790029049, + 0.0007874165312387049, + 0.002149431500583887, + 0.0012374406214803457, + 0.0022390212398022413, + 0.0032380581833422184, + 0.0021200706250965595, + 0.0017519916873425245, + 0.0032842939253896475, + 0.0032074416521936655, + 0.003559848526492715, + 0.0055172438733279705, + 0.003080197377130389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.029787329956889153 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 lush, tall grass with individually rendered blades featuring soft edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the grass blades glowing with vibrant greens and golds under warm sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute smooth, low-angle sideways trajectories gliding through the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a forgotten metal kettle lying on its side, partially covered in reddish-brown rust, visible in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a tiny ladybug with a polished crimson shell and black spot crawling along the curve of the kettle's spout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978559397161007 + ], + "opposite_logprob_tokens": [ + -3.529785633087158 + ], + "logprob_reward": -0.02978559397161007, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978559397161007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038786791265010834, + 0.014223022386431694, + 0.01706654205918312, + 0.15067173540592194, + 0.01196630671620369, + 0.021555142477154732, + 0.6961343288421631, + 0.013054732233285904, + 0.014263499528169632, + 0.022277818992733955 + ], + "credit_map": [ + 0.0035586142912507057, + 0.0034493342973291874, + 0.008363458327949047, + 0.0049085114151239395, + 0.0034818127751350403, + 0.006988194305449724, + 0.0032724139746278524, + 0.004764449782669544, + 0.001154193072579801, + 0.0009159155888482928, + 0.003601440927013755, + 0.0022185181733220816, + 0.0012534206034615636, + 0.0017481037648394704, + 0.001692792633548379, + 0.0016386377392336726, + 0.0026856844779103994, + 0.0031074320431798697, + 0.0019409266533330083, + 0.0012607964454218745, + 0.0014944234862923622, + 0.0018301032250747085, + 0.0014194418909028172, + 0.0033277347683906555, + 0.010887141339480877, + 0.05674581602215767, + 0.004216752015054226, + 0.0036453770007938147, + 0.006998246535658836, + 0.05577399209141731, + 0.0034209999721497297, + 0.008983410894870758, + 0.0013336186530068517, + 0.0014748949324712157, + 0.0017800607020035386, + 0.0014584894524887204, + 0.0011272052070125937, + 0.0016277822433039546, + 0.001609979197382927, + 0.0015542764449492097, + 0.002139598596841097, + 0.0015884970780462027, + 0.0019340675789862871, + 0.0022825361229479313, + 0.0014987415634095669, + 0.004595749080181122, + 0.003992793150246143, + 0.0035231595393270254, + 0.07874956727027893, + 0.011179055087268353, + 0.02111678756773472, + 0.012363661080598831, + 0.02675657719373703, + 0.3067100942134857, + 0.03619671240448952, + 0.20306186378002167, + 0.0014334951993077993, + 0.0011575574753805995, + 0.0012604979565367103, + 0.0017179371789097786, + 0.0012580747716128826, + 0.002256209496408701, + 0.0022955862805247307, + 0.0016753729432821274, + 0.0020117645617574453, + 0.0008353546145372093, + 0.0014298328896984458, + 0.001877660513855517, + 0.000994491740129888, + 0.002200442599132657, + 0.0027856389060616493, + 0.0021283128298819065, + 0.0021514941472560167, + 0.001702524023130536, + 0.0013028006069362164, + 0.00331737264059484, + 0.0024197339080274105, + 0.0027778921648859978, + 0.006109909620136023, + 0.0024960918817669153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.02978559397161007 + } + ], + "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/0000110/group_5/index_tensor([1896, 1896])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_5/index_tensor([1896, 1896])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..643aab55cb68184be1f8ae937f10256788a9a474 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_5/index_tensor([1896, 1896])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping long shot, shot from a slightly upward angle, pans slowly in an arc around the center of the battle, revealing a mythical duel taking place atop a swirling sea of clouds. A massive dragon, covered in lush green scales resembling jade, its slender, sinuous body like a moving mountain, tightly entwines a snow-white tiger of equal size with its powerful body. Each scale of the dragon gleams with a blend of metallic and gemstone in the light filtering through the clouds, its golden pupils blazing with fury. The tiger, imprisoned by it, is a pure white, without a trace of other colors. Its fur, hard as steel from the throes of its muscles, its icy blue eyes brimming with savagery and murderous intent. The tiger struggles fiercely, its sharp claws digging deep into the cracks in the dragon's scales, attempting to tear through its sturdy defenses. At the same time, its bloody maw opens, revealing dagger-like fangs, as it bites fiercely at the dragon's neck. Their fierce struggle churns the dense clouds, forming massive vortexes and turbulence. Golden sunlight penetrates the gaps in the clouds, casting magnificent rays of light (the Tyndall effect), precisely illuminating the flying scale fragments and scattered white tiger fur in the air caused by the collision. The entire scene is filled with a sense of primitive, wild, and sacred power, creating a tense and magnificent epic atmosphere.", + "seed": 21230, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_5/index_1896_seed_21230_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": -0.032472206326201555, + "vlm_logprob": -0.032472206326201555, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.032472206326201555, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005309646483510733 + ], + "opposite_logprob_tokens": [ + -6.005309581756592 + ], + "logprob_reward": -0.005309646483510733, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005309646483510733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5198736190795898, + 0.030223019421100616, + 0.018265845254063606, + 0.018205642700195312, + 0.03989870473742485, + 0.0560469888150692, + 0.09845994412899017, + 0.07711216062307358, + 0.10560545325279236, + 0.036308642476797104 + ], + "credit_map": [ + 0.057558536529541016, + 0.0469987615942955, + 0.06075391173362732, + 0.05097610875964165, + 0.05945737659931183, + 0.0784902274608612, + 0.08788568526506424, + 0.0777529776096344, + 0.0035673801321536303, + 0.0030648319516330957, + 0.006942396052181721, + 0.0027486952021718025, + 0.003075777553021908, + 0.003400034736841917, + 0.003357120556756854, + 0.004066782537847757, + 0.0020321044139564037, + 0.0023829913698136806, + 0.0024846827145665884, + 0.0018663491355255246, + 0.0026170930359512568, + 0.002017408376559615, + 0.0024773499462753534, + 0.0023878663778305054, + 0.002044126857072115, + 0.0019081479404121637, + 0.002768375677987933, + 0.0021740735974162817, + 0.00181104545481503, + 0.002189493039622903, + 0.0027175270952284336, + 0.0025928535033017397, + 0.002717840252444148, + 0.0023255115374922752, + 0.009549091570079327, + 0.00484986649826169, + 0.0026203300803899765, + 0.007230513729155064, + 0.0033461791463196278, + 0.007259368430823088, + 0.004885120317339897, + 0.004780332092195749, + 0.00909600593149662, + 0.004098589066416025, + 0.004149097017943859, + 0.008260488510131836, + 0.009478447027504444, + 0.011298910714685917, + 0.013277661986649036, + 0.00955404993146658, + 0.009367634542286396, + 0.009171218611299992, + 0.006604422349482775, + 0.015604780055582523, + 0.01701296679675579, + 0.017867211252450943, + 0.005824204068630934, + 0.0034928221721202135, + 0.006203121040016413, + 0.006369732785969973, + 0.005175504367798567, + 0.014259656891226768, + 0.01225027535110712, + 0.023536844179034233, + 0.007372633554041386, + 0.015108908526599407, + 0.015266593545675278, + 0.011038919910788536, + 0.009413929656147957, + 0.01569179818034172, + 0.009830727241933346, + 0.0218819472938776, + 0.003339912509545684, + 0.004654458723962307, + 0.004434480797499418, + 0.004469086416065693, + 0.009551146999001503, + 0.003120762761682272, + 0.0031201306264847517, + 0.003618664341047406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.005309646483510733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002832450671121478 + ], + "opposite_logprob_tokens": [ + -5.877832412719727 + ], + "logprob_reward": -0.002832450671121478, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002832450671121478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24666178226470947, + 0.050669506192207336, + 0.06701655685901642, + 0.06496382504701614, + 0.05293309688568115, + 0.02865847758948803, + 0.0947234109044075, + 0.034827083349227905, + 0.3080499470233917, + 0.05149632692337036 + ], + "credit_map": [ + 0.025671323761343956, + 0.029183482751250267, + 0.02413150481879711, + 0.02285781130194664, + 0.02906535007059574, + 0.036641042679548264, + 0.042808789759874344, + 0.036302484571933746, + 0.004229886922985315, + 0.006153750699013472, + 0.008939303457736969, + 0.003742215456441045, + 0.006936124991625547, + 0.0070909131318330765, + 0.0069290464743971825, + 0.006648267153650522, + 0.006992391776293516, + 0.011082956567406654, + 0.008647286333143711, + 0.008142379112541676, + 0.01107226312160492, + 0.0063357301987707615, + 0.006102551240473986, + 0.008641001768410206, + 0.006285474635660648, + 0.0073913708329200745, + 0.010116668418049812, + 0.008138095028698444, + 0.006035518832504749, + 0.007940556854009628, + 0.010112041607499123, + 0.008944100700318813, + 0.004383798222988844, + 0.00450794817879796, + 0.01111641712486744, + 0.0070445481687784195, + 0.003934939857572317, + 0.008476758375763893, + 0.0054586962796747684, + 0.008009997196495533, + 0.0019115757895633578, + 0.002140951808542013, + 0.003741611260920763, + 0.002047301270067692, + 0.0024779431987553835, + 0.004921140614897013, + 0.006631491705775261, + 0.0047864606603980064, + 0.012168669141829014, + 0.011433745734393597, + 0.01039494015276432, + 0.007575095631182194, + 0.007639417890459299, + 0.014571836218237877, + 0.012538173235952854, + 0.018401537090539932, + 0.0027412099298089743, + 0.002519382396712899, + 0.003155649406835437, + 0.003425877308472991, + 0.0023602647706866264, + 0.006253884639590979, + 0.0037460867315530777, + 0.010624728165566921, + 0.024241099134087563, + 0.033269815146923065, + 0.037750907242298126, + 0.03494291380047798, + 0.03487241268157959, + 0.04526912420988083, + 0.034102000296115875, + 0.06360168009996414, + 0.004989484325051308, + 0.006357800215482712, + 0.006322207860648632, + 0.0064346217550337315, + 0.011477119289338589, + 0.004508974030613899, + 0.004884159192442894, + 0.006521960720419884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.002832450671121478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076287128031253815 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.0076287128031253815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076287128031253815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3092036247253418, + 0.05440651625394821, + 0.027314171195030212, + 0.03551208972930908, + 0.040279991924762726, + 0.059814050793647766, + 0.0889192447066307, + 0.016964934766292572, + 0.3327547311782837, + 0.03483065217733383 + ], + "credit_map": [ + 0.035383451730012894, + 0.028689930215477943, + 0.03483128547668457, + 0.03131963312625885, + 0.033118899911642075, + 0.04638848081231117, + 0.05371551588177681, + 0.045756444334983826, + 0.0058250101283192635, + 0.004626919515430927, + 0.007676037028431892, + 0.006561261136084795, + 0.00910049956291914, + 0.006783342454582453, + 0.007691531907767057, + 0.0061419131234288216, + 0.002805026713758707, + 0.004022256471216679, + 0.003586355596780777, + 0.003548622829839587, + 0.004151108209043741, + 0.0029852865263819695, + 0.0027334787882864475, + 0.003482036292552948, + 0.0035038902424275875, + 0.004260279703885317, + 0.005443674977868795, + 0.004443329758942127, + 0.0032589351758360863, + 0.004808453377336264, + 0.0044153006747365, + 0.005378228146582842, + 0.0021511130034923553, + 0.002494141226634383, + 0.009685344994068146, + 0.004958942532539368, + 0.0021292301826179028, + 0.007906116545200348, + 0.003378312336280942, + 0.007576789241284132, + 0.0075271339155733585, + 0.006729776971042156, + 0.007148194592446089, + 0.004041401669383049, + 0.00615670345723629, + 0.008346960879862309, + 0.008632084354758263, + 0.011231792159378529, + 0.010235504247248173, + 0.006188297178596258, + 0.0039079450070858, + 0.004754751455038786, + 0.004047834314405918, + 0.018593965098261833, + 0.016833169385790825, + 0.02435777708888054, + 0.0015408974140882492, + 0.0015522035537287593, + 0.002541877329349518, + 0.002147720195353031, + 0.0017086765728890896, + 0.0022327560000121593, + 0.0016395049169659615, + 0.0036012986674904823, + 0.02608591504395008, + 0.038548439741134644, + 0.042167194187641144, + 0.03813692554831505, + 0.03717474639415741, + 0.04724742844700813, + 0.036519818007946014, + 0.06687424331903458, + 0.003908465150743723, + 0.0037544614169746637, + 0.005074838642030954, + 0.004241287242621183, + 0.006879700347781181, + 0.0029006069526076317, + 0.003597359638661146, + 0.004473929759114981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0076287128031253815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290998756885529 + ], + "opposite_logprob_tokens": [ + -2.237910032272339 + ], + "logprob_reward": -0.11290998756885529, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290998756885529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4326772391796112, + 0.06444326043128967, + 0.05955389142036438, + 0.1745326668024063, + 0.037652816623449326, + 0.027580469846725464, + 0.05205249413847923, + 0.023449061438441277, + 0.09697946906089783, + 0.031078657135367393 + ], + "credit_map": [ + 0.04930443316698074, + 0.0388651005923748, + 0.04630494490265846, + 0.042700909078121185, + 0.0452478863298893, + 0.06376189738512039, + 0.07841582596302032, + 0.06807620823383331, + 0.005332782864570618, + 0.006723536644130945, + 0.011914115399122238, + 0.005820041056722403, + 0.009021105244755745, + 0.008449663408100605, + 0.008370856754481792, + 0.008811158128082752, + 0.006285968702286482, + 0.008295905776321888, + 0.008381390012800694, + 0.0065331365913152695, + 0.008131432346999645, + 0.006620718166232109, + 0.006691196002066135, + 0.008614139631390572, + 0.015157670713961124, + 0.017751626670360565, + 0.027775539085268974, + 0.023834211751818657, + 0.015099422074854374, + 0.023023778572678566, + 0.027068158611655235, + 0.024822253733873367, + 0.0024415806401520967, + 0.002833710517734289, + 0.007827766239643097, + 0.007163966540247202, + 0.0025677732191979885, + 0.005850627087056637, + 0.004175559151917696, + 0.004791831132024527, + 0.002134781563654542, + 0.001663835602812469, + 0.004051859024912119, + 0.0022441905457526445, + 0.0015904833562672138, + 0.004792830906808376, + 0.005952662788331509, + 0.005149825941771269, + 0.005640303250402212, + 0.005436633713543415, + 0.004830438178032637, + 0.003711224300786853, + 0.003982963040471077, + 0.008671383373439312, + 0.00713855167850852, + 0.012640991248190403, + 0.0027657051105052233, + 0.0016904979711398482, + 0.0023641425650566816, + 0.0023410888388752937, + 0.0017836384940892458, + 0.0038625290617346764, + 0.0023861515801399946, + 0.00625530444085598, + 0.007433820515871048, + 0.013622201047837734, + 0.011600637808442116, + 0.009295111522078514, + 0.008485028520226479, + 0.014989707618951797, + 0.008090893737971783, + 0.02346206083893776, + 0.003235872834920883, + 0.0038674992974847555, + 0.004339348990470171, + 0.00387184857390821, + 0.006170508451759815, + 0.002534713363274932, + 0.0029985099099576473, + 0.004060354083776474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.11290998756885529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368023410439491 + ], + "opposite_logprob_tokens": [ + -3.4086802005767822 + ], + "logprob_reward": -0.03368023410439491, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368023410439491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22005636990070343, + 0.06722629070281982, + 0.1179138571023941, + 0.08159563690423965, + 0.07144077122211456, + 0.1554470956325531, + 0.03503657504916191, + 0.025679394602775574, + 0.10094209760427475, + 0.12466197460889816 + ], + "credit_map": [ + 0.021676888689398766, + 0.020652610808610916, + 0.02635144256055355, + 0.019417179748415947, + 0.02533641830086708, + 0.03633300960063934, + 0.037800710648298264, + 0.032488081604242325, + 0.006451633758842945, + 0.0076349275186657906, + 0.009759697131812572, + 0.007390771061182022, + 0.011050920933485031, + 0.00847566593438387, + 0.00846506655216217, + 0.007997602224349976, + 0.01793387532234192, + 0.014674529433250427, + 0.017205823212862015, + 0.014010950922966003, + 0.016758419573307037, + 0.012003830634057522, + 0.01134475227445364, + 0.013981667347252369, + 0.006287284195423126, + 0.007205583155155182, + 0.014746109955012798, + 0.009342820383608341, + 0.007156165782362223, + 0.012765000574290752, + 0.01023951917886734, + 0.013853141106665134, + 0.004010822158306837, + 0.004644285887479782, + 0.017855407670140266, + 0.007907469756901264, + 0.0042291060090065, + 0.013073020614683628, + 0.0073041464202106, + 0.012416502460837364, + 0.017545290291309357, + 0.018365640193223953, + 0.01974503882229328, + 0.01323915645480156, + 0.016605272889137268, + 0.021392498165369034, + 0.020290156826376915, + 0.02826402708888054, + 0.0027023975271731615, + 0.0034230859018862247, + 0.0027536950074136257, + 0.0030664235819131136, + 0.0029882255475968122, + 0.005523153115063906, + 0.005798782221972942, + 0.008780808188021183, + 0.0022822332102805376, + 0.002361689694225788, + 0.0026350142434239388, + 0.0025003713089972734, + 0.0020304671488702297, + 0.004285897593945265, + 0.002552354009822011, + 0.0070313639007508755, + 0.008315718732774258, + 0.016398832201957703, + 0.01262685190886259, + 0.008567304350435734, + 0.012201722711324692, + 0.013401619158685207, + 0.009431353770196438, + 0.019998684525489807, + 0.011642981320619583, + 0.012127391993999481, + 0.019369374960660934, + 0.014053157530725002, + 0.01988561451435089, + 0.012828596867620945, + 0.014016002416610718, + 0.020738841965794563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.03368023410439491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, shot from a slightly upward angle, pans slowly in an arc around the center of the battle, revealing a mythical duel taking place atop a swirling sea of clouds. A massive dragon, covered in lush green scales resembling jade, its slender, sinuous body like a moving mountain, tightly entwines a snow-white tiger of equal size with its powerful body. Each scale of the dragon gleams with a blend of metallic and gemstone in the light filtering through the clouds, its golden pupils blazing with fury. The tiger, imprisoned by it, is a pure white, without a trace of other colors. Its fur, hard as steel from the throes of its muscles, its icy blue eyes brimming with savagery and murderous intent. The tiger struggles fiercely, its sharp claws digging deep into the cracks in the dragon's scales, attempting to tear through its sturdy defenses. At the same time, its bloody maw opens, revealing dagger-like fangs, as it bites fiercely at the dragon's neck. Their fierce struggle churns the dense clouds, forming massive vortexes and turbulence. Golden sunlight penetrates the gaps in the clouds, casting magnificent rays of light (the Tyndall effect), precisely illuminating the flying scale fragments and scattered white tiger fur in the air caused by the collision. The entire scene is filled with a sense of primitive, wild, and sacred power, creating a tense and magnificent epic atmosphere.", + "seed": 21231, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_5/index_1896_seed_21231_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": -0.09262965125963092, + "vlm_logprob": -0.09262965125963092, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.09262965125963092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006969309411942959 + ], + "opposite_logprob_tokens": [ + -5.256969451904297 + ], + "logprob_reward": -0.006969309411942959, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006969309411942959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10636012256145477, + 0.09232842177152634, + 0.10660859942436218, + 0.0747540295124054, + 0.07751164585351944, + 0.07382069528102875, + 0.23279742896556854, + 0.033857561647892, + 0.1243007630109787, + 0.07766072452068329 + ], + "credit_map": [ + 0.007571620866656303, + 0.008012205362319946, + 0.014117629267275333, + 0.011705717071890831, + 0.014936404302716255, + 0.023561935871839523, + 0.010554355569183826, + 0.015900244936347008, + 0.006257811561226845, + 0.009800540283322334, + 0.0076657854951918125, + 0.010669511742889881, + 0.009074129164218903, + 0.014584246091544628, + 0.01426851749420166, + 0.020007874816656113, + 0.008285999298095703, + 0.012958275154232979, + 0.008458459749817848, + 0.018560927361249924, + 0.01240751426666975, + 0.015170122496783733, + 0.019980715587735176, + 0.01078658364713192, + 0.006896547507494688, + 0.011039618402719498, + 0.005131790414452553, + 0.008603000082075596, + 0.015184246003627777, + 0.01060890406370163, + 0.008356026373803616, + 0.00893389992415905, + 0.006291160825639963, + 0.010603143833577633, + 0.008720061741769314, + 0.009968745522201061, + 0.015283582732081413, + 0.009682457894086838, + 0.008611018769443035, + 0.00835147500038147, + 0.0043256343342363834, + 0.004108094610273838, + 0.010289521887898445, + 0.006815212778747082, + 0.005369885824620724, + 0.01228068396449089, + 0.01281492318958044, + 0.017816737294197083, + 0.018329838290810585, + 0.01924889162182808, + 0.036131251603364944, + 0.025101102888584137, + 0.03109128773212433, + 0.026090532541275024, + 0.029601559042930603, + 0.04720296338200569, + 0.003466174006462097, + 0.0025009112432599068, + 0.005326834507286549, + 0.006015442777425051, + 0.0033827638253569603, + 0.0032171227503567934, + 0.005190834403038025, + 0.004757476504892111, + 0.012939275242388248, + 0.009116126224398613, + 0.011318976059556007, + 0.025562401860952377, + 0.010416761972010136, + 0.012755690142512321, + 0.026968939229846, + 0.01522259321063757, + 0.009016650728881359, + 0.008848385885357857, + 0.013779137283563614, + 0.009671292267739773, + 0.009427711367607117, + 0.009846560657024384, + 0.0096671674400568, + 0.00740381795912981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.006969309411942959 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036245384253561497 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.0036245384253561497, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036245384253561497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09403515607118607, + 0.03535601124167442, + 0.2326481193304062, + 0.08548104017972946, + 0.03543517366051674, + 0.04427563026547432, + 0.05380130931735039, + 0.08995623141527176, + 0.1127294972538948, + 0.21628183126449585 + ], + "credit_map": [ + 0.005471629556268454, + 0.006178585346788168, + 0.008186236023902893, + 0.009597978554666042, + 0.009110442362725735, + 0.02452753856778145, + 0.011249272152781487, + 0.019713453948497772, + 0.003824562067165971, + 0.0037157831247895956, + 0.002577228005975485, + 0.006629229057580233, + 0.0030820192769169807, + 0.004664561711251736, + 0.004748406819999218, + 0.006114220712333918, + 0.021720027551054955, + 0.03211995214223862, + 0.024986565113067627, + 0.027085034176707268, + 0.027473989874124527, + 0.031246304512023926, + 0.046195030212402344, + 0.021821198984980583, + 0.008234667591750622, + 0.014004004187881947, + 0.004713618196547031, + 0.010192866437137127, + 0.01802079752087593, + 0.013079920783638954, + 0.006716144736856222, + 0.010519013740122318, + 0.0032215153332799673, + 0.005428968463093042, + 0.004170376807451248, + 0.004081241320818663, + 0.005486542358994484, + 0.004532493185251951, + 0.0041268072091042995, + 0.004387227352708578, + 0.0023982874117791653, + 0.003146107541397214, + 0.006213619373738766, + 0.0056991237215697765, + 0.0038471848238259554, + 0.006772071123123169, + 0.00977130513638258, + 0.006427926477044821, + 0.004471552092581987, + 0.004029903095215559, + 0.008643905632197857, + 0.005370423197746277, + 0.008163152262568474, + 0.008450611494481564, + 0.006807474419474602, + 0.007864286191761494, + 0.010318982414901257, + 0.006294178310781717, + 0.01469874195754528, + 0.017158152535557747, + 0.0077978624030947685, + 0.010404136963188648, + 0.011530875228345394, + 0.011753288097679615, + 0.008126034401357174, + 0.0073278481140732765, + 0.011485228314995766, + 0.026735249906778336, + 0.00955786183476448, + 0.00714252470061183, + 0.02958749607205391, + 0.012767250649631023, + 0.025071462616324425, + 0.02645113505423069, + 0.031087137758731842, + 0.02692251466214657, + 0.03713409602642059, + 0.02465670369565487, + 0.02489433065056801, + 0.02006443217396736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.0036245384253561497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076287128031253815 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.0076287128031253815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076287128031253815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08470940589904785, + 0.061904653906822205, + 0.117937371134758, + 0.07872667908668518, + 0.1000652089715004, + 0.044545359909534454, + 0.0500631108880043, + 0.10736878216266632, + 0.214365616440773, + 0.1403137892484665 + ], + "credit_map": [ + 0.006113236770033836, + 0.00657055526971817, + 0.009565388783812523, + 0.010642618872225285, + 0.006611470598727465, + 0.019821258261799812, + 0.008959167636930943, + 0.016425715759396553, + 0.004235167521983385, + 0.005665038712322712, + 0.0042961458675563335, + 0.01272828970104456, + 0.005854611285030842, + 0.008368127048015594, + 0.008760522119700909, + 0.011996755376458168, + 0.010031100362539291, + 0.015886802226305008, + 0.008481285534799099, + 0.017594199627637863, + 0.015090280212461948, + 0.0158951785415411, + 0.022695263847708702, + 0.012263267301023006, + 0.009509745985269547, + 0.01454212050884962, + 0.004969676956534386, + 0.011970213614404202, + 0.012899483554065228, + 0.010665726847946644, + 0.005441487766802311, + 0.008728227578103542, + 0.006939247716218233, + 0.008847824297845364, + 0.00899419654160738, + 0.007048811297863722, + 0.024177853018045425, + 0.019239988178014755, + 0.008408607915043831, + 0.016408687457442284, + 0.0028544904198497534, + 0.003343154676258564, + 0.005320601165294647, + 0.005497705191373825, + 0.003660377347841859, + 0.00685541657730937, + 0.010524031706154346, + 0.006489585619419813, + 0.0037759453989565372, + 0.0033688524272292852, + 0.007596323266625404, + 0.005933794658631086, + 0.005840363912284374, + 0.006683671846985817, + 0.007314465008676052, + 0.009549698792397976, + 0.011994398199021816, + 0.007757491432130337, + 0.017488079145550728, + 0.01864197477698326, + 0.009792461059987545, + 0.012526866979897022, + 0.01422622799873352, + 0.014941294677555561, + 0.016246644780039787, + 0.01184110064059496, + 0.019602598622441292, + 0.05475345626473427, + 0.015778664499521255, + 0.011960672214627266, + 0.058310575783252716, + 0.02587190829217434, + 0.016170868650078773, + 0.0172744058072567, + 0.019401732832193375, + 0.01829138770699501, + 0.024376576766371727, + 0.01579786278307438, + 0.016345703974366188, + 0.012655258178710938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.0076287128031253815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288647174835205 + ], + "opposite_logprob_tokens": [ + -1.0538647174835205 + ], + "logprob_reward": -0.4288647174835205, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288647174835205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08909803628921509, + 0.05398266762495041, + 0.06706993281841278, + 0.10949314385652542, + 0.06931200623512268, + 0.034575603902339935, + 0.049929749220609665, + 0.1303105652332306, + 0.29099881649017334, + 0.10522949695587158 + ], + "credit_map": [ + 0.0057061901316046715, + 0.0070706866681575775, + 0.007758110761642456, + 0.008560911752283573, + 0.009029324166476727, + 0.022008785977959633, + 0.010145856067538261, + 0.018818166106939316, + 0.004369599744677544, + 0.006740884855389595, + 0.004525009077042341, + 0.0075345830991864204, + 0.006895702797919512, + 0.006532955914735794, + 0.00875796377658844, + 0.008625967428088188, + 0.005363031756132841, + 0.0067800735123455524, + 0.00496299983933568, + 0.007570796646177769, + 0.007725409232079983, + 0.010549916885793209, + 0.017340095713734627, + 0.006777608301490545, + 0.010229279287159443, + 0.020013421773910522, + 0.006087743677198887, + 0.01305289939045906, + 0.024266380816698074, + 0.013974696397781372, + 0.008612089790403843, + 0.013256626203656197, + 0.004701647907495499, + 0.009254218079149723, + 0.008183750323951244, + 0.008032063022255898, + 0.010946099646389484, + 0.01127282902598381, + 0.007408568169921637, + 0.009512832388281822, + 0.0037928379606455564, + 0.002814582549035549, + 0.004264933057129383, + 0.003728964366018772, + 0.0022743477020412683, + 0.004978662356734276, + 0.007076808717101812, + 0.00564446859061718, + 0.004046621732413769, + 0.0035384439397603273, + 0.008439465425908566, + 0.005713288672268391, + 0.00531799579039216, + 0.006994863972067833, + 0.007332885637879372, + 0.008546181954443455, + 0.014027110300958157, + 0.008915113285183907, + 0.022110359743237495, + 0.022838324308395386, + 0.010581587441265583, + 0.01557980477809906, + 0.017287882044911385, + 0.018970385193824768, + 0.02034742571413517, + 0.016431357711553574, + 0.027064206078648567, + 0.07508678734302521, + 0.02172532118856907, + 0.015757160261273384, + 0.08073043823242188, + 0.033856138586997986, + 0.01188488770276308, + 0.012934694066643715, + 0.014649380929768085, + 0.013326535001397133, + 0.018381018191576004, + 0.012203892692923546, + 0.012099693529307842, + 0.009749396704137325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.4288647174835205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060978174209595 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.016060978174209595, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060978174209595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15869751572608948, + 0.1609184592962265, + 0.052215591073036194, + 0.15226168930530548, + 0.08748038858175278, + 0.06659003347158432, + 0.05370061472058296, + 0.06647887080907822, + 0.0861208438873291, + 0.11553601175546646 + ], + "credit_map": [ + 0.013450098223984241, + 0.009023364633321762, + 0.015947073698043823, + 0.015527314506471157, + 0.019974425435066223, + 0.03934192657470703, + 0.016679003834724426, + 0.028754306957125664, + 0.015856647863984108, + 0.022456210106611252, + 0.011855006217956543, + 0.026536429300904274, + 0.022359874099493027, + 0.01798110269010067, + 0.01900223083794117, + 0.024870960041880608, + 0.004493548069149256, + 0.006849117577075958, + 0.003329600440338254, + 0.008861929178237915, + 0.008124292828142643, + 0.006753568071871996, + 0.007371508050709963, + 0.006432026624679565, + 0.014616595581173897, + 0.029933009296655655, + 0.008840713649988174, + 0.01825486496090889, + 0.02750377729535103, + 0.022334717214107513, + 0.012934602797031403, + 0.017843399196863174, + 0.007451443001627922, + 0.014952931553125381, + 0.010839567519724369, + 0.010472559370100498, + 0.013179227709770203, + 0.009804405272006989, + 0.010339911095798016, + 0.010440346784889698, + 0.0056585087440907955, + 0.006339834071695805, + 0.011316575109958649, + 0.008499735966324806, + 0.00593635393306613, + 0.009314856491982937, + 0.010567354038357735, + 0.00895681232213974, + 0.003997350577265024, + 0.004535633604973555, + 0.008162645623087883, + 0.006623345427215099, + 0.005259241908788681, + 0.0088938158005476, + 0.007487209979444742, + 0.008741372264921665, + 0.006857785861939192, + 0.005099981091916561, + 0.010987884365022182, + 0.009778098203241825, + 0.007180013228207827, + 0.007425300311297178, + 0.009456002153456211, + 0.0096938107162714, + 0.009270296432077885, + 0.006160603370517492, + 0.013992746360599995, + 0.015298831276595592, + 0.007612344343215227, + 0.008003992028534412, + 0.016189560294151306, + 0.009592468850314617, + 0.014806536026299, + 0.01665455847978592, + 0.013667205348610878, + 0.01545946765691042, + 0.023251470178365707, + 0.01184024941176176, + 0.010586187243461609, + 0.009270338341593742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.016060978174209595 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, shot from a slightly upward angle, pans slowly in an arc around the center of the battle, revealing a mythical duel taking place atop a swirling sea of clouds. A massive dragon, covered in lush green scales resembling jade, its slender, sinuous body like a moving mountain, tightly entwines a snow-white tiger of equal size with its powerful body. Each scale of the dragon gleams with a blend of metallic and gemstone in the light filtering through the clouds, its golden pupils blazing with fury. The tiger, imprisoned by it, is a pure white, without a trace of other colors. Its fur, hard as steel from the throes of its muscles, its icy blue eyes brimming with savagery and murderous intent. The tiger struggles fiercely, its sharp claws digging deep into the cracks in the dragon's scales, attempting to tear through its sturdy defenses. At the same time, its bloody maw opens, revealing dagger-like fangs, as it bites fiercely at the dragon's neck. Their fierce struggle churns the dense clouds, forming massive vortexes and turbulence. Golden sunlight penetrates the gaps in the clouds, casting magnificent rays of light (the Tyndall effect), precisely illuminating the flying scale fragments and scattered white tiger fur in the air caused by the collision. The entire scene is filled with a sense of primitive, wild, and sacred power, creating a tense and magnificent epic atmosphere.", + "seed": 21232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_5/index_1896_seed_21232_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": -0.02517241782043129, + "vlm_logprob": -0.02517241782043129, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02517241782043129, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013580316677689552 + ], + "opposite_logprob_tokens": [ + -4.388580322265625 + ], + "logprob_reward": -0.013580316677689552, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013580316677689552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21465525031089783, + 0.07459473609924316, + 0.09001941978931427, + 0.04606255888938904, + 0.06713414192199707, + 0.034691669046878815, + 0.09557152539491653, + 0.16409985721111298, + 0.09697888046503067, + 0.11619190871715546 + ], + "credit_map": [ + 0.012335489504039288, + 0.01837204210460186, + 0.021546248346567154, + 0.015748558565974236, + 0.01829308271408081, + 0.0561385452747345, + 0.031483255326747894, + 0.04073803499341011, + 0.00519636832177639, + 0.010663485154509544, + 0.013241126202046871, + 0.010699592530727386, + 0.0067305006086826324, + 0.00894076842814684, + 0.0063628084026277065, + 0.012760087847709656, + 0.011113738641142845, + 0.006806214805692434, + 0.022263916209340096, + 0.005446915049105883, + 0.0048919073306024075, + 0.018062986433506012, + 0.006201829761266708, + 0.015231906436383724, + 0.003930232487618923, + 0.004771342501044273, + 0.010396292433142662, + 0.0030707877594977617, + 0.004075290635228157, + 0.005268862470984459, + 0.004842281341552734, + 0.009707469493150711, + 0.00514299888163805, + 0.008130457252264023, + 0.004815451800823212, + 0.005258641671389341, + 0.00845745112746954, + 0.011777250096201897, + 0.006474155467003584, + 0.017077738419175148, + 0.0034488928504288197, + 0.0035214528907090425, + 0.0038822642527520657, + 0.003956375177949667, + 0.004520696122199297, + 0.002741309581324458, + 0.004281473346054554, + 0.00833920855075121, + 0.009678271599113941, + 0.01151058729737997, + 0.018389105796813965, + 0.008951741270720959, + 0.011226480826735497, + 0.017888633534312248, + 0.007140117231756449, + 0.010786585509777069, + 0.024432726204395294, + 0.016653098165988922, + 0.017983736470341682, + 0.012779110111296177, + 0.01628817990422249, + 0.019089194014668465, + 0.01829744502902031, + 0.038576364517211914, + 0.009828900918364525, + 0.0053289299830794334, + 0.006222106982022524, + 0.005392096936702728, + 0.008908779360353947, + 0.031399067491292953, + 0.007828295230865479, + 0.02207069844007492, + 0.028938643634319305, + 0.00888000801205635, + 0.036593589931726456, + 0.006848218385130167, + 0.007429084274917841, + 0.007410801947116852, + 0.014290266670286655, + 0.005801287479698658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.013580316677689552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019481267081573606 + ], + "opposite_logprob_tokens": [ + -6.251948356628418 + ], + "logprob_reward": -0.0019481267081573606, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019481267081573606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13694509863853455, + 0.04617881774902344, + 0.052813105285167694, + 0.06143425032496452, + 0.2787426710128784, + 0.0945412814617157, + 0.07310511916875839, + 0.06262432783842087, + 0.10499756783246994, + 0.0886177346110344 + ], + "credit_map": [ + 0.0104343481361866, + 0.010888857766985893, + 0.01656975783407688, + 0.015148629434406757, + 0.015271716751158237, + 0.023879095911979675, + 0.025235868990421295, + 0.019516831263899803, + 0.00365111930295825, + 0.003948488738387823, + 0.007937299087643623, + 0.008042230270802975, + 0.0029555123765021563, + 0.005705916788429022, + 0.005302002653479576, + 0.008636252023279667, + 0.004312779754400253, + 0.0062477136962115765, + 0.00754268653690815, + 0.005664119031280279, + 0.004840869456529617, + 0.010911297984421253, + 0.005794420372694731, + 0.007499223109334707, + 0.00804776418954134, + 0.00718408590182662, + 0.010049327276647091, + 0.004697594791650772, + 0.010074658319354057, + 0.007173504214733839, + 0.00831624772399664, + 0.005891073495149612, + 0.028758520260453224, + 0.035669587552547455, + 0.024853069335222244, + 0.02352878823876381, + 0.0314004085958004, + 0.032230813056230545, + 0.03135688975453377, + 0.07094460725784302, + 0.008151785470545292, + 0.005062723997980356, + 0.021552514284849167, + 0.0066001322120428085, + 0.007666950114071369, + 0.007782415021210909, + 0.007210663985460997, + 0.030514096841216087, + 0.009472079575061798, + 0.009435953572392464, + 0.01307037752121687, + 0.009241314604878426, + 0.009867900982499123, + 0.009305731393396854, + 0.005816077347844839, + 0.00689568929374218, + 0.008079981431365013, + 0.006337405648082495, + 0.005970275495201349, + 0.008094560354948044, + 0.009599595330655575, + 0.007831576280295849, + 0.009358599781990051, + 0.007352335378527641, + 0.006319153588265181, + 0.013960451819002628, + 0.0063733686693012714, + 0.008756478317081928, + 0.026828452944755554, + 0.017834076657891273, + 0.010829598642885685, + 0.014095999300479889, + 0.02581283636391163, + 0.009590838104486465, + 0.01526616420596838, + 0.006036465056240559, + 0.008158472366631031, + 0.006608211435377598, + 0.007521744817495346, + 0.009623009711503983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0019481267081573606 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251185968518257 + ], + "opposite_logprob_tokens": [ + -5.255251407623291 + ], + "logprob_reward": -0.005251185968518257, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251185968518257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1677471399307251, + 0.07477130740880966, + 0.04676475748419762, + 0.03436841070652008, + 0.33161690831184387, + 0.027578171342611313, + 0.04693794995546341, + 0.09978614002466202, + 0.10544297099113464, + 0.06498623639345169 + ], + "credit_map": [ + 0.011667841114103794, + 0.011922686360776424, + 0.017242901027202606, + 0.016918938606977463, + 0.014745551161468029, + 0.042976345866918564, + 0.022738443687558174, + 0.029534440487623215, + 0.005019571166485548, + 0.006488012615591288, + 0.013449818827211857, + 0.00860859639942646, + 0.0049403696320950985, + 0.008959711529314518, + 0.007963616400957108, + 0.019341617822647095, + 0.00315576558932662, + 0.003784777596592903, + 0.010639077983796597, + 0.0044275918044149876, + 0.00443774089217186, + 0.008875778876245022, + 0.004923527128994465, + 0.006520501337945461, + 0.0055723111145198345, + 0.004321885760873556, + 0.005003251135349274, + 0.0029830732382833958, + 0.004358784761279821, + 0.003534747753292322, + 0.0037781656719744205, + 0.004816190805286169, + 0.04612334072589874, + 0.06624343246221542, + 0.026167457923293114, + 0.03662104159593582, + 0.047010526061058044, + 0.027002517133951187, + 0.04070854187011719, + 0.041740067303180695, + 0.0028993459418416023, + 0.0031647777650505304, + 0.0030016994569450617, + 0.003099260851740837, + 0.00398108409717679, + 0.0023140888661146164, + 0.0031496123410761356, + 0.005968303885310888, + 0.008104636333882809, + 0.004509050864726305, + 0.007352928165346384, + 0.004837497603148222, + 0.0046449401415884495, + 0.008879137225449085, + 0.0035538075026124716, + 0.005055955145508051, + 0.009173024445772171, + 0.012204890139400959, + 0.010573129169642925, + 0.014628013595938683, + 0.015011955052614212, + 0.009192119352519512, + 0.018740616738796234, + 0.010262400843203068, + 0.009957381524145603, + 0.009659123606979847, + 0.006885590963065624, + 0.008960614912211895, + 0.014914056286215782, + 0.025533201172947884, + 0.009558138437569141, + 0.019974878057837486, + 0.01683923415839672, + 0.005513035226613283, + 0.015463078394532204, + 0.003945660777390003, + 0.0050443378277122974, + 0.005743986461311579, + 0.007628640625625849, + 0.004808271303772926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.005251185968518257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901972323656082 + ], + "opposite_logprob_tokens": [ + -2.464019775390625 + ], + "logprob_reward": -0.08901972323656082, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901972323656082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12889952957630157, + 0.05592942237854004, + 0.028830263763666153, + 0.06517508625984192, + 0.09519635885953903, + 0.021938586607575417, + 0.06857573986053467, + 0.047102365642786026, + 0.09042127430438995, + 0.39793139696121216 + ], + "credit_map": [ + 0.007935742847621441, + 0.010766993276774883, + 0.010771457105875015, + 0.008848629891872406, + 0.008936061523854733, + 0.04173321649432182, + 0.02788383699953556, + 0.012023581191897392, + 0.003803903702646494, + 0.004682753700762987, + 0.009616667404770851, + 0.00876166857779026, + 0.0038270947989076376, + 0.006170073989778757, + 0.006963534280657768, + 0.012103731743991375, + 0.0028000432066619396, + 0.003385959193110466, + 0.002972615882754326, + 0.00407568272203207, + 0.003544643521308899, + 0.003983532078564167, + 0.003706555347889662, + 0.004361229948699474, + 0.012056425213813782, + 0.006241085473448038, + 0.012814524583518505, + 0.0050280336290597916, + 0.005925092380493879, + 0.00671018660068512, + 0.007943256758153439, + 0.00845648068934679, + 0.009451936930418015, + 0.008788066916167736, + 0.007354418747127056, + 0.008607327938079834, + 0.008391566574573517, + 0.015656976029276848, + 0.011383079923689365, + 0.025562986731529236, + 0.0021975699346512556, + 0.00227434653788805, + 0.0023019113577902317, + 0.002075513591989875, + 0.002564124995842576, + 0.002011287957429886, + 0.0028075366280972958, + 0.005706296768039465, + 0.007927021011710167, + 0.009227335453033447, + 0.012554258108139038, + 0.005880127660930157, + 0.007964018732309341, + 0.011687771417200565, + 0.005178909283131361, + 0.008156299591064453, + 0.004055233206599951, + 0.007629530970007181, + 0.0049627372063696384, + 0.005864105653017759, + 0.007506645284593105, + 0.004570272751152515, + 0.007553248666226864, + 0.004960592836141586, + 0.008724954910576344, + 0.00841943547129631, + 0.005133775062859058, + 0.0052317064255476, + 0.012359037064015865, + 0.024763528257608414, + 0.008134507574141026, + 0.01765432581305504, + 0.10003767907619476, + 0.032259199768304825, + 0.11737596988677979, + 0.026265723630785942, + 0.029443345963954926, + 0.027651669457554817, + 0.047219082713127136, + 0.017678700387477875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.08901972323656082 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606273651123047 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.01606273651123047, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606273651123047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12077783793210983, + 0.05650964751839638, + 0.051432035863399506, + 0.07042396068572998, + 0.06797808408737183, + 0.05271531641483307, + 0.10647702217102051, + 0.06289909034967422, + 0.12139041721820831, + 0.28939658403396606 + ], + "credit_map": [ + 0.008067905902862549, + 0.007369496859610081, + 0.014018756337463856, + 0.008424543775618076, + 0.007590447552502155, + 0.03273995965719223, + 0.02621087059378624, + 0.016355860978364944, + 0.003905121237039566, + 0.004192539025098085, + 0.010448119603097439, + 0.007861093617975712, + 0.0040421076118946075, + 0.007756776642054319, + 0.005828518886119127, + 0.01247536949813366, + 0.005244934931397438, + 0.0035364963114261627, + 0.010310456156730652, + 0.0034514078870415688, + 0.004241117741912603, + 0.010798796080052853, + 0.005927719175815582, + 0.007921110838651657, + 0.008954294957220554, + 0.008578057400882244, + 0.013436184264719486, + 0.006483655888587236, + 0.008169193752110004, + 0.007708808407187462, + 0.008907297626137733, + 0.008186469785869122, + 0.005086467135697603, + 0.008139687590301037, + 0.00960423145443201, + 0.005279205739498138, + 0.010812249034643173, + 0.005339056719094515, + 0.006517759058624506, + 0.017199425026774406, + 0.005993509199470282, + 0.004781501833349466, + 0.009325703606009483, + 0.0048484462313354015, + 0.005590109620243311, + 0.0054816692136228085, + 0.005162276793271303, + 0.011532100848853588, + 0.010058244690299034, + 0.014049852266907692, + 0.021739574149250984, + 0.009293694980442524, + 0.012961730360984802, + 0.019039029255509377, + 0.007686023600399494, + 0.0116488728672266, + 0.008920222520828247, + 0.008935194462537766, + 0.005858118645846844, + 0.006960366852581501, + 0.0067757535725831985, + 0.008058274164795876, + 0.009381992742419243, + 0.00800916738808155, + 0.009906229563057423, + 0.010901153087615967, + 0.009764106012880802, + 0.00918674562126398, + 0.01403780747205019, + 0.032508864998817444, + 0.010914355516433716, + 0.024171151220798492, + 0.08420214802026749, + 0.02168651856482029, + 0.08526783436536789, + 0.01687340810894966, + 0.017438525334000587, + 0.01773579604923725, + 0.03225002810359001, + 0.013942340388894081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.01606273651123047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping long shot, shot from a slightly upward angle, pans slowly in an arc around the center of the battle, revealing a mythical duel taking place atop a swirling sea of clouds. A massive dragon, covered in lush green scales resembling jade, its slender, sinuous body like a moving mountain, tightly entwines a snow-white tiger of equal size with its powerful body. Each scale of the dragon gleams with a blend of metallic and gemstone in the light filtering through the clouds, its golden pupils blazing with fury. The tiger, imprisoned by it, is a pure white, without a trace of other colors. Its fur, hard as steel from the throes of its muscles, its icy blue eyes brimming with savagery and murderous intent. The tiger struggles fiercely, its sharp claws digging deep into the cracks in the dragon's scales, attempting to tear through its sturdy defenses. At the same time, its bloody maw opens, revealing dagger-like fangs, as it bites fiercely at the dragon's neck. Their fierce struggle churns the dense clouds, forming massive vortexes and turbulence. Golden sunlight penetrates the gaps in the clouds, casting magnificent rays of light (the Tyndall effect), precisely illuminating the flying scale fragments and scattered white tiger fur in the air caused by the collision. The entire scene is filled with a sense of primitive, wild, and sacred power, creating a tense and magnificent epic atmosphere.", + "seed": 21233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_5/index_1896_seed_21233_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": -0.04084210586734116, + "vlm_logprob": -0.04084210586734116, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.04084210586734116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020128745585680008 + ], + "opposite_logprob_tokens": [ + -4.0201287269592285 + ], + "logprob_reward": -0.020128745585680008, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.020128745585680008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16699406504631042, + 0.04356859251856804, + 0.02898488938808441, + 0.039902474731206894, + 0.05473640188574791, + 0.1430286318063736, + 0.02905457094311714, + 0.32396674156188965, + 0.14463934302330017, + 0.02512429468333721 + ], + "credit_map": [ + 0.03102085553109646, + 0.021889066323637962, + 0.018459724262356758, + 0.01837359555065632, + 0.023415686562657356, + 0.020861495286226273, + 0.016335319727659225, + 0.016638318076729774, + 0.005904058460146189, + 0.0036515924148261547, + 0.008759546093642712, + 0.003491005627438426, + 0.0073658437468111515, + 0.005740572698414326, + 0.0034053116105496883, + 0.005250666756182909, + 0.003651436883956194, + 0.0036823975387960672, + 0.004684868734329939, + 0.0026106948498636484, + 0.0036384453997015953, + 0.0037653048057109118, + 0.003436172381043434, + 0.0035155676305294037, + 0.004048189148306847, + 0.0030624214559793472, + 0.005249677691608667, + 0.0027516582049429417, + 0.004650391172617674, + 0.010824304074048996, + 0.0027540400624275208, + 0.006561792455613613, + 0.007375152315944433, + 0.007251877337694168, + 0.00552317313849926, + 0.00510781304910779, + 0.0054280501790344715, + 0.01065338309854269, + 0.00519806332886219, + 0.008198888972401619, + 0.02008245512843132, + 0.01280881091952324, + 0.01753111183643341, + 0.0134299136698246, + 0.01513406541198492, + 0.028473662212491035, + 0.017281468957662582, + 0.01828714646399021, + 0.004081561230123043, + 0.0036242089699953794, + 0.0024928513448685408, + 0.0034962890204042196, + 0.0029933294281363487, + 0.003520834492519498, + 0.0038186185993254185, + 0.00502688018605113, + 0.030181536450982094, + 0.028413988649845123, + 0.03486127406358719, + 0.0278669074177742, + 0.054582759737968445, + 0.06924203783273697, + 0.04017531871795654, + 0.038642894476652145, + 0.012786954641342163, + 0.01096539106220007, + 0.01726892963051796, + 0.015154012478888035, + 0.008201087825000286, + 0.05327731370925903, + 0.011521417647600174, + 0.01546422764658928, + 0.0030000980477780104, + 0.003697575768455863, + 0.004433894529938698, + 0.003232498886063695, + 0.002520371228456497, + 0.0029698158614337444, + 0.0021855528466403484, + 0.00308448844589293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.020128745585680008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001726209418848157 + ], + "opposite_logprob_tokens": [ + -6.376726150512695 + ], + "logprob_reward": -0.001726209418848157, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001726209418848157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17177443206310272, + 0.09410706907510757, + 0.10354216396808624, + 0.04866435378789902, + 0.04064175486564636, + 0.06391863524913788, + 0.09338963031768799, + 0.13485586643218994, + 0.20505501329898834, + 0.04405105859041214 + ], + "credit_map": [ + 0.03346867486834526, + 0.018412990495562553, + 0.02273656241595745, + 0.013573432341217995, + 0.0270173791795969, + 0.021397091448307037, + 0.01362715195864439, + 0.021541167050600052, + 0.009692494757473469, + 0.008001379668712616, + 0.010549901984632015, + 0.00749612832441926, + 0.01694987341761589, + 0.01940067857503891, + 0.008042579516768456, + 0.013974041678011417, + 0.01064245868474245, + 0.015223653987050056, + 0.015313594602048397, + 0.008256329223513603, + 0.013873650692403316, + 0.016215480864048004, + 0.011119398288428783, + 0.012897616252303123, + 0.005609807558357716, + 0.00800062995404005, + 0.0050012776628136635, + 0.005048729479312897, + 0.007286803796887398, + 0.007287974469363689, + 0.0052408999763429165, + 0.005188239272683859, + 0.008031393401324749, + 0.0056159356608986855, + 0.00473526306450367, + 0.004331330768764019, + 0.003970571793615818, + 0.00456838496029377, + 0.004392193164676428, + 0.004996687173843384, + 0.010444077663123608, + 0.009701305069029331, + 0.006393594667315483, + 0.008779422380030155, + 0.006751161068677902, + 0.008327709510922432, + 0.006419215351343155, + 0.007102157920598984, + 0.024218613281846046, + 0.008764105848968029, + 0.007275799755007029, + 0.006619441322982311, + 0.008568757213652134, + 0.0096665658056736, + 0.013490145094692707, + 0.014786209911108017, + 0.019973454996943474, + 0.01300128549337387, + 0.014712691307067871, + 0.014174677431583405, + 0.017541199922561646, + 0.019881514832377434, + 0.022428052499890327, + 0.013143005780875683, + 0.015796123072504997, + 0.01961405575275421, + 0.02841177210211754, + 0.03811751678586006, + 0.01484114583581686, + 0.032834552228450775, + 0.02108139917254448, + 0.034358467906713486, + 0.006472861859947443, + 0.00618885550647974, + 0.007159993518143892, + 0.005066241137683392, + 0.0033958791755139828, + 0.005099206231534481, + 0.005139073822647333, + 0.005528953392058611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.001726209418848157 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952251143753529 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952251143753529, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952251143753529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19712954759597778, + 0.10631468147039413, + 0.06666485220193863, + 0.027278654277324677, + 0.05605992302298546, + 0.07303401827812195, + 0.0491563118994236, + 0.09502597898244858, + 0.24865858256816864, + 0.08067742735147476 + ], + "credit_map": [ + 0.03896816074848175, + 0.02361195534467697, + 0.02394711785018444, + 0.016188912093639374, + 0.031174223870038986, + 0.020973069593310356, + 0.017993319779634476, + 0.024272799491882324, + 0.009074869565665722, + 0.01124877855181694, + 0.007198309525847435, + 0.009427395649254322, + 0.01745733432471752, + 0.02436211332678795, + 0.011058424599468708, + 0.016487469896674156, + 0.00795749295502901, + 0.009080599062144756, + 0.012377067469060421, + 0.00554690370336175, + 0.009764115326106548, + 0.007166207302361727, + 0.006336756516247988, + 0.008435718715190887, + 0.004627195652574301, + 0.0035651251673698425, + 0.0035800545010715723, + 0.0030779456719756126, + 0.0038447135593742132, + 0.002897944301366806, + 0.0027516221161931753, + 0.002934057032689452, + 0.009024661034345627, + 0.008505051955580711, + 0.005167525727301836, + 0.005807965993881226, + 0.005539913661777973, + 0.008717740885913372, + 0.005434349644929171, + 0.007862718775868416, + 0.01135457307100296, + 0.012900114059448242, + 0.006485727149993181, + 0.008960428647696972, + 0.00875807460397482, + 0.009342527948319912, + 0.007554175797849894, + 0.007678409572690725, + 0.008811167441308498, + 0.00710460776463151, + 0.004566794261336327, + 0.004552265163511038, + 0.004293245263397694, + 0.004989273846149445, + 0.006084966007620096, + 0.008753996342420578, + 0.009111017920076847, + 0.008582030422985554, + 0.011483562178909779, + 0.011458820663392544, + 0.01273658312857151, + 0.009972921572625637, + 0.018355920910835266, + 0.013325128704309464, + 0.023937750607728958, + 0.027988485991954803, + 0.029425514861941338, + 0.04541020467877388, + 0.021639080718159676, + 0.041900258511304855, + 0.017609458416700363, + 0.040747836232185364, + 0.009198388084769249, + 0.011284979060292244, + 0.013330789282917976, + 0.01147516630589962, + 0.009096493944525719, + 0.008334572426974773, + 0.00996214710175991, + 0.007994895800948143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.005952251143753529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034211218357086 + ], + "opposite_logprob_tokens": [ + -1.9103420972824097 + ], + "logprob_reward": -0.16034211218357086, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034211218357086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660573124885559, + 0.06480005383491516, + 0.025506196543574333, + 0.0413762703537941, + 0.04919293522834778, + 0.06640772521495819, + 0.15339599549770355, + 0.04763262718915939, + 0.13807323575019836, + 0.14755764603614807 + ], + "credit_map": [ + 0.05836230888962746, + 0.024906925857067108, + 0.03244843706488609, + 0.0278252474963665, + 0.03825848922133446, + 0.025345202535390854, + 0.024766702204942703, + 0.034143995493650436, + 0.006036197766661644, + 0.007493580691516399, + 0.006040474865585566, + 0.007550682872533798, + 0.010097143240272999, + 0.010985570028424263, + 0.007375084795057774, + 0.009221323765814304, + 0.0027508616913110018, + 0.002988178515806794, + 0.004072331357747316, + 0.002642480656504631, + 0.003993263002485037, + 0.0027776379138231277, + 0.002876683371141553, + 0.003404762828722596, + 0.00660084979608655, + 0.005741155240684748, + 0.005275439005345106, + 0.004953288938850164, + 0.006610351148992777, + 0.003953465260565281, + 0.0038527906872332096, + 0.004388932604342699, + 0.008956418372690678, + 0.007095567882061005, + 0.004470422398298979, + 0.00551508367061615, + 0.0046113706193864346, + 0.007036964409053326, + 0.004816240631043911, + 0.006690868176519871, + 0.009521663188934326, + 0.009278474375605583, + 0.006176856812089682, + 0.0069692726247012615, + 0.008941864594817162, + 0.009864709340035915, + 0.007951784878969193, + 0.0077031077817082405, + 0.04610076919198036, + 0.016074175015091896, + 0.009129419922828674, + 0.010470454581081867, + 0.010115866549313068, + 0.011571657843887806, + 0.008587365038692951, + 0.04134628176689148, + 0.0055112093687057495, + 0.004452703054994345, + 0.006200205534696579, + 0.005700165405869484, + 0.004426281433552504, + 0.006943576969206333, + 0.008029102347791195, + 0.006369387730956078, + 0.0131652457639575, + 0.01614764705300331, + 0.019685544073581696, + 0.02673860266804695, + 0.01062880177050829, + 0.01623154804110527, + 0.010677124373614788, + 0.024798721075057983, + 0.016927707940340042, + 0.024297136813402176, + 0.021051043644547462, + 0.018329255282878876, + 0.01669440045952797, + 0.015683477744460106, + 0.012483605183660984, + 0.022091027349233627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.16034211218357086 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 atop a swirling sea of clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive dragon with lush green jade-like scales visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a snow-white tiger with icy blue eyes visible, tightly entwined with the dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters engaged in a fierce struggle where the tiger bites the dragon's neck and claws dig into its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan slowly in an arc around the center of the battle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606121100485325 + ], + "opposite_logprob_tokens": [ + -4.141061305999756 + ], + "logprob_reward": -0.01606121100485325, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606121100485325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10425782203674316, + 0.052182942628860474, + 0.029961995780467987, + 0.04595791548490524, + 0.08234070241451263, + 0.06035787984728813, + 0.09579198062419891, + 0.17163021862506866, + 0.3280178904533386, + 0.029500557109713554 + ], + "credit_map": [ + 0.020770153030753136, + 0.014602365903556347, + 0.011778819374740124, + 0.00828530639410019, + 0.013826719485223293, + 0.013444552198052406, + 0.009599154815077782, + 0.011950744315981865, + 0.005475263576954603, + 0.0032678372226655483, + 0.00500964792445302, + 0.0029111537151038647, + 0.009621069766581059, + 0.014668785966932774, + 0.0033622058108448982, + 0.007866974920034409, + 0.002876260317862034, + 0.0023106366861611605, + 0.004203490447252989, + 0.0031648569274693727, + 0.004572586622089148, + 0.005264284089207649, + 0.0037339148111641407, + 0.0038359682075679302, + 0.004780604504048824, + 0.006526704411953688, + 0.00574538204818964, + 0.004354660399258137, + 0.007636534050107002, + 0.005747468210756779, + 0.004389029927551746, + 0.00677752960473299, + 0.009852305985987186, + 0.016130203381180763, + 0.00774862477555871, + 0.010236873291432858, + 0.008004835806787014, + 0.011708807200193405, + 0.0079030217602849, + 0.01075602788478136, + 0.013203962706029415, + 0.0062609342858195305, + 0.008781002834439278, + 0.005395801737904549, + 0.005124824121594429, + 0.006657387129962444, + 0.006627128459513187, + 0.008306843228638172, + 0.023164838552474976, + 0.01468460913747549, + 0.008683789521455765, + 0.008556712418794632, + 0.007725830189883709, + 0.005755699705332518, + 0.013291019015014172, + 0.013929483480751514, + 0.025739453732967377, + 0.01818583719432354, + 0.018180422484874725, + 0.02214874140918255, + 0.013266509398818016, + 0.01903543993830681, + 0.03427857160568237, + 0.020795252174139023, + 0.03182462975382805, + 0.03749953582882881, + 0.046998947858810425, + 0.06190091371536255, + 0.025858059525489807, + 0.044754091650247574, + 0.03045053966343403, + 0.048731185495853424, + 0.003341514151543379, + 0.004655071534216404, + 0.004729905165731907, + 0.003381676971912384, + 0.003285893239080906, + 0.0030862174462527037, + 0.002737839240580797, + 0.004282438196241856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.01606121100485325 + } + ], + "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/0000110/group_6/index_tensor([5595, 5595])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_tensor([5595, 5595])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..afff099c5d13c7584d62281325debc7f8bda0b95 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_tensor([5595, 5595])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low, fixed wide-angle shot, shot from above, captures the moment when the swirling, inky clouds of the heavens are suddenly ripped apart by an invisible force. The Monkey King, adorned with an intricately carved, majestically shimmering phoenix-wing purple-gold crown, draped in golden chain mail armor, and tied with an ochre-yellow tiger-skin skirt, bursts explosively from beneath the clouds. His face is defiant, his eyes wide open, radiating a piercing golden light. His bared teeth utter a silent war cry. He clutches the radiant golden Ruyi Jingu Bang in both hands and swings it forward with thunderous force. The energy halo surrounding the club instantly pierces the thick wall of cloud, sending countless shattered clouds into a churning wave that scatters in all directions, forming a vast, ever-expanding vortex-like void behind him. The whole picture is full of epic power and the magnificence of breaking free from constraints. The golden sacred light forms a sharp contrast with the dark clouds, creating a mythical, grand and fierce battle atmosphere.", + "seed": 3658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_5595_seed_3658_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": -0.15697687175124883, + "vlm_logprob": -0.15697687175124883, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.15697687175124883, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7034175395965576 + ], + "opposite_logprob_tokens": [ + -0.7034175395965576 + ], + "logprob_reward": -0.7034175395965576, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7034175395965576, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2551303207874298, + 0.08060204237699509, + 0.03557403385639191, + 0.4099864959716797, + 0.04044130817055702, + 0.044299256056547165, + 0.047715507447719574, + 0.03038470633327961, + 0.01826387643814087, + 0.03760240972042084 + ], + "credit_map": [ + 0.017990175634622574, + 0.028026629239320755, + 0.023866916075348854, + 0.03393847495317459, + 0.04189081862568855, + 0.050266001373529434, + 0.034628648310899734, + 0.024522656574845314, + 0.0062940060161054134, + 0.007509144023060799, + 0.018755540251731873, + 0.01047209370881319, + 0.005695310886949301, + 0.0051419297233223915, + 0.0068316832184791565, + 0.019902337342500687, + 0.002869516611099243, + 0.004614611156284809, + 0.005875681526958942, + 0.005358920432627201, + 0.0028534953016787767, + 0.003234818112105131, + 0.005883901380002499, + 0.00488309096544981, + 0.11575023829936981, + 0.023000501096248627, + 0.045931071043014526, + 0.024485161527991295, + 0.018037553876638412, + 0.10411893576383591, + 0.015306220389902592, + 0.06335680931806564, + 0.004781682044267654, + 0.005745018366724253, + 0.0029620893765240908, + 0.006510483101010323, + 0.005165728274732828, + 0.004719375632703304, + 0.0055203791707754135, + 0.005036551039665937, + 0.0035771254915744066, + 0.006310288328677416, + 0.003098777262493968, + 0.005946811754256487, + 0.004786524921655655, + 0.007130161859095097, + 0.005816902499645948, + 0.0076326634734869, + 0.005307378713041544, + 0.007004183251410723, + 0.004445068538188934, + 0.006823343224823475, + 0.00418788893148303, + 0.006694389041513205, + 0.006762007251381874, + 0.00649124663323164, + 0.003143936162814498, + 0.0037627711426466703, + 0.0024056686088442802, + 0.0051673767156898975, + 0.0029043960385024548, + 0.0036579223815351725, + 0.004633864853531122, + 0.004708771128207445, + 0.0019045366207137704, + 0.002751346444711089, + 0.0017977042589336634, + 0.0024575195275247097, + 0.0017751578707247972, + 0.00302514317445457, + 0.0022203074768185616, + 0.002332160249352455, + 0.004670977126806974, + 0.003544494742527604, + 0.005728177260607481, + 0.004403416067361832, + 0.004863885696977377, + 0.004511111415922642, + 0.004582812078297138, + 0.005297533702105284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.7034175395965576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563489377498627 + ], + "opposite_logprob_tokens": [ + -4.387563705444336 + ], + "logprob_reward": -0.012563489377498627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563489377498627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10495023429393768, + 0.05276598036289215, + 0.05233648791909218, + 0.37296947836875916, + 0.06665509939193726, + 0.06514977663755417, + 0.035566218197345734, + 0.1330922394990921, + 0.0600418858230114, + 0.05647264048457146 + ], + "credit_map": [ + 0.007701203227043152, + 0.010619902051985264, + 0.01042897067964077, + 0.02510601095855236, + 0.010403652675449848, + 0.014907388016581535, + 0.015100698918104172, + 0.010682410560548306, + 0.004381696693599224, + 0.005674885585904121, + 0.01034074928611517, + 0.007297666743397713, + 0.005090244114398956, + 0.005778992548584938, + 0.004974760580807924, + 0.009226981550455093, + 0.0053225611336529255, + 0.00832268875092268, + 0.007162866182625294, + 0.007321259938180447, + 0.005332439672201872, + 0.0047210692428052425, + 0.007713494822382927, + 0.006440109573304653, + 0.11518751084804535, + 0.027424870058894157, + 0.06159259006381035, + 0.02463129721581936, + 0.017432820051908493, + 0.07081668823957443, + 0.014556227251887321, + 0.04132744297385216, + 0.006488428451120853, + 0.010313172824680805, + 0.004386437125504017, + 0.01120795402675867, + 0.006534320302307606, + 0.00980012770742178, + 0.007513535674661398, + 0.010411120019853115, + 0.0064294603653252125, + 0.0077871885150671005, + 0.004998828750103712, + 0.00944734737277031, + 0.007065030746161938, + 0.008561749942600727, + 0.010924628004431725, + 0.009935539215803146, + 0.00393930496647954, + 0.005783966742455959, + 0.0032565391156822443, + 0.004420108627527952, + 0.0033581985626369715, + 0.005132464226335287, + 0.004400522913783789, + 0.005275109317153692, + 0.0139546487480402, + 0.021014897152781487, + 0.010754667222499847, + 0.029348639771342278, + 0.013922542333602905, + 0.012668298557400703, + 0.01904813013970852, + 0.012380408123135567, + 0.007578657474368811, + 0.008046192117035389, + 0.00437961146235466, + 0.00791641604155302, + 0.0042196777649223804, + 0.008627516217529774, + 0.007766021881252527, + 0.011507794260978699, + 0.0049964324571192265, + 0.00853725429624319, + 0.005252808798104525, + 0.010181321762502193, + 0.004100812133401632, + 0.005389668047428131, + 0.010587724857032299, + 0.007426618132740259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.012563489377498627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2180926501750946, + 0.051887936890125275, + 0.06463438272476196, + 0.21767501533031464, + 0.2013116478919983, + 0.07310963422060013, + 0.04927553981542587, + 0.04983993619680405, + 0.05273304507136345, + 0.021440111100673676 + ], + "credit_map": [ + 0.01600927859544754, + 0.016975192353129387, + 0.02267855778336525, + 0.04638483747839928, + 0.024496031925082207, + 0.02403206378221512, + 0.04679485782980919, + 0.020721841603517532, + 0.004076996352523565, + 0.009008674882352352, + 0.006279695313423872, + 0.009375769644975662, + 0.005707136821001768, + 0.0042173778638243675, + 0.006567809730768204, + 0.006654480006545782, + 0.007831976749002934, + 0.009480150416493416, + 0.008199595846235752, + 0.007549993693828583, + 0.0074103460647165775, + 0.006192705128341913, + 0.009433620609343052, + 0.008535996079444885, + 0.05949040874838829, + 0.0221013855189085, + 0.035441868007183075, + 0.01869671791791916, + 0.016046367585659027, + 0.03280862420797348, + 0.01252948772162199, + 0.020560186356306076, + 0.01695432886481285, + 0.02087530307471752, + 0.008377564139664173, + 0.0177567508071661, + 0.014501267112791538, + 0.05203871428966522, + 0.015522066503763199, + 0.05528568848967552, + 0.01182241179049015, + 0.006358979735523462, + 0.006760825868695974, + 0.00810721330344677, + 0.005936693400144577, + 0.010388931259512901, + 0.009886126965284348, + 0.013848458416759968, + 0.00556351151317358, + 0.009293491020798683, + 0.0028258543461561203, + 0.0079137422144413, + 0.0046511488035321236, + 0.0074180676601827145, + 0.005310066044330597, + 0.006299660541117191, + 0.004277474712580442, + 0.006376356817781925, + 0.0040879263542592525, + 0.009968274272978306, + 0.004985913168638945, + 0.005804788786917925, + 0.006408389192074537, + 0.007930818013846874, + 0.005916420370340347, + 0.005015087313950062, + 0.0036046982277184725, + 0.004454205743968487, + 0.0028409913647919893, + 0.011756791733205318, + 0.005002582911401987, + 0.014142269268631935, + 0.0021078824065625668, + 0.0032224601600319147, + 0.0017081331461668015, + 0.003655638312920928, + 0.0016826244536787271, + 0.0020465778652578592, + 0.005270898807793856, + 0.0017458995571359992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.016072357073426247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305015504360199 + ], + "opposite_logprob_tokens": [ + -3.1680500507354736 + ], + "logprob_reward": -0.04305015504360199, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305015504360199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13375061750411987, + 0.04309593141078949, + 0.20418627560138702, + 0.13604356348514557, + 0.19417577981948853, + 0.11748325079679489, + 0.04290224239230156, + 0.06189500913023949, + 0.039074696600437164, + 0.027392631396651268 + ], + "credit_map": [ + 0.01143705751746893, + 0.011171706020832062, + 0.015175419859588146, + 0.01932811178267002, + 0.024300826713442802, + 0.02670726738870144, + 0.011848454363644123, + 0.013781781308352947, + 0.0035184479784220457, + 0.007466753479093313, + 0.006074118427932262, + 0.007565921172499657, + 0.0048204748891294, + 0.003559042466804385, + 0.005074524786323309, + 0.005016650073230267, + 0.017295239493250847, + 0.030266785994172096, + 0.027628198266029358, + 0.030194388702511787, + 0.02361595630645752, + 0.017958855256438255, + 0.03022470325231552, + 0.02700214460492134, + 0.04301734268665314, + 0.011741885915398598, + 0.02625495009124279, + 0.010140041820704937, + 0.008986745029687881, + 0.017741771414875984, + 0.007256439886987209, + 0.010904398746788502, + 0.014868239872157574, + 0.02062130905687809, + 0.007014923263341188, + 0.015730788931250572, + 0.013138258829712868, + 0.053221482783555984, + 0.01303113717585802, + 0.05654966086149216, + 0.016165807843208313, + 0.012357916682958603, + 0.010663661174476147, + 0.013195730745792389, + 0.011103962548077106, + 0.016264528036117554, + 0.016119971871376038, + 0.02161167375743389, + 0.003921836148947477, + 0.006344430148601532, + 0.0039212871342897415, + 0.008538298308849335, + 0.0036392344627529383, + 0.004813783802092075, + 0.006815948523581028, + 0.004907427821308374, + 0.00669611943885684, + 0.010154059156775475, + 0.006547144148498774, + 0.011319268494844437, + 0.007024759892374277, + 0.005815480835735798, + 0.008680450730025768, + 0.005657731089740992, + 0.008431474678218365, + 0.0032238296698778868, + 0.0031335626263171434, + 0.0030067581683397293, + 0.0023057605139911175, + 0.00537702115252614, + 0.003073823871091008, + 0.010522466152906418, + 0.002666252665221691, + 0.006044410634785891, + 0.0019974035676568747, + 0.004595646634697914, + 0.0019334802636876702, + 0.002460705116391182, + 0.004994478076696396, + 0.002700256183743477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.04305015504360199 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1615065485239029, + 0.1173582598567009, + 0.13371163606643677, + 0.3117164075374603, + 0.05933696776628494, + 0.048474255949258804, + 0.03205983713269234, + 0.04928365722298622, + 0.04470231011509895, + 0.04185009375214577 + ], + "credit_map": [ + 0.014590906910598278, + 0.0185565035790205, + 0.020952902734279633, + 0.029782351106405258, + 0.016820406541228294, + 0.020190035924315453, + 0.02302221953868866, + 0.017591208219528198, + 0.008778621442615986, + 0.010446528904139996, + 0.022803349420428276, + 0.015431473962962627, + 0.012202218174934387, + 0.011105680838227272, + 0.011931744404137135, + 0.02465863898396492, + 0.010771457105875015, + 0.016690053045749664, + 0.019351493567228317, + 0.022686058655381203, + 0.012566142715513706, + 0.00976639799773693, + 0.024293702095746994, + 0.01758633740246296, + 0.08443912118673325, + 0.01865268312394619, + 0.03211028501391411, + 0.019529862329363823, + 0.017443614080548286, + 0.07844985276460648, + 0.014810617081820965, + 0.04628035053610802, + 0.005740117281675339, + 0.007921112701296806, + 0.003350235056132078, + 0.008581635542213917, + 0.0057989973574876785, + 0.010616693645715714, + 0.006736512761563063, + 0.010591665282845497, + 0.004871435463428497, + 0.005799256265163422, + 0.004500797018408775, + 0.006779090501368046, + 0.004270502831786871, + 0.006696820259094238, + 0.008077388629317284, + 0.0074789635837078094, + 0.0028123415540903807, + 0.0047493972815573215, + 0.004121704492717981, + 0.0037076082080602646, + 0.002510339254513383, + 0.004727741237729788, + 0.003765510395169258, + 0.005665196571499109, + 0.005608151201158762, + 0.005836927331984043, + 0.004782629664987326, + 0.00761822797358036, + 0.006298532709479332, + 0.005313328001648188, + 0.006403093226253986, + 0.007422765251249075, + 0.004719840828329325, + 0.007849003188312054, + 0.004013443365693092, + 0.005229724105447531, + 0.005794479977339506, + 0.00641766469925642, + 0.005273978225886822, + 0.005404177121818066, + 0.0044363997876644135, + 0.007995734922587872, + 0.0034637716598808765, + 0.005411830265074968, + 0.0033167467918246984, + 0.003877056995406747, + 0.008232605643570423, + 0.005115947220474482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.009780817665159702 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, fixed wide-angle shot, shot from above, captures the moment when the swirling, inky clouds of the heavens are suddenly ripped apart by an invisible force. The Monkey King, adorned with an intricately carved, majestically shimmering phoenix-wing purple-gold crown, draped in golden chain mail armor, and tied with an ochre-yellow tiger-skin skirt, bursts explosively from beneath the clouds. His face is defiant, his eyes wide open, radiating a piercing golden light. His bared teeth utter a silent war cry. He clutches the radiant golden Ruyi Jingu Bang in both hands and swings it forward with thunderous force. The energy halo surrounding the club instantly pierces the thick wall of cloud, sending countless shattered clouds into a churning wave that scatters in all directions, forming a vast, ever-expanding vortex-like void behind him. The whole picture is full of epic power and the magnificence of breaking free from constraints. The golden sacred light forms a sharp contrast with the dark clouds, creating a mythical, grand and fierce battle atmosphere.", + "seed": 3659, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_5595_seed_3659_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": -0.14047389244660735, + "vlm_logprob": -0.14047389244660735, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.14047389244660735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6404001712799072 + ], + "opposite_logprob_tokens": [ + -0.7654001712799072 + ], + "logprob_reward": -0.6404001712799072, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6404001712799072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13043048977851868, + 0.03537809103727341, + 0.0948105901479721, + 0.08010045439004898, + 0.10129895806312561, + 0.2239185869693756, + 0.11798524856567383, + 0.11229332536458969, + 0.05954110622406006, + 0.044243115931749344 + ], + "credit_map": [ + 0.010304152965545654, + 0.01897040382027626, + 0.018605411052703857, + 0.014736557379364967, + 0.014984780922532082, + 0.02682461030781269, + 0.012858572416007519, + 0.01314601767808199, + 0.00458552548661828, + 0.004118444863706827, + 0.0039671240374445915, + 0.0031763953156769276, + 0.005070452578365803, + 0.003180112922564149, + 0.0061315083876252174, + 0.0051485332660377026, + 0.008267590776085854, + 0.0038494605105370283, + 0.031748849898576736, + 0.00878128968179226, + 0.00487899873405695, + 0.0057412683963775635, + 0.005713649094104767, + 0.025829488411545753, + 0.01107579842209816, + 0.008025208488106728, + 0.00930271577090025, + 0.007985454984009266, + 0.009748924523591995, + 0.012133431620895863, + 0.008682850748300552, + 0.013146079145371914, + 0.009356815367937088, + 0.008092974312603474, + 0.019556351006031036, + 0.004918246995657682, + 0.010803546756505966, + 0.027484292164444923, + 0.004678842145949602, + 0.01640789955854416, + 0.009349620901048183, + 0.007025460246950388, + 0.0884653627872467, + 0.00674401456490159, + 0.009800940752029419, + 0.07558484375476837, + 0.004253844730556011, + 0.022694533690810204, + 0.01704453118145466, + 0.011836707592010498, + 0.013660651631653309, + 0.011928373947739601, + 0.01847483217716217, + 0.011570778675377369, + 0.018166949972510338, + 0.015302441082894802, + 0.013209386728703976, + 0.008446111343801022, + 0.012814482674002647, + 0.01366092637181282, + 0.01614033617079258, + 0.012331194244325161, + 0.025200925767421722, + 0.01048996951431036, + 0.007661640178412199, + 0.005504267290234566, + 0.01013987883925438, + 0.007186968810856342, + 0.005144897382706404, + 0.010194024071097374, + 0.00500087533146143, + 0.008708560839295387, + 0.008061413653194904, + 0.003888334147632122, + 0.012536942027509212, + 0.004684544634073973, + 0.0027744730468839407, + 0.005995818879455328, + 0.0027925751637667418, + 0.003509021131321788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.6404001712799072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033724840730428696 + ], + "opposite_logprob_tokens": [ + -3.408724784851074 + ], + "logprob_reward": -0.033724840730428696, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033724840730428696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11860594153404236, + 0.028838403522968292, + 0.06049644201993942, + 0.11535084992647171, + 0.10820625722408295, + 0.2959948778152466, + 0.037055954337120056, + 0.15171775221824646, + 0.026324806734919548, + 0.057408690452575684 + ], + "credit_map": [ + 0.010762125253677368, + 0.010560083203017712, + 0.012980648316442966, + 0.01465521939098835, + 0.011253530159592628, + 0.018012376502156258, + 0.02526591159403324, + 0.015116065740585327, + 0.003349070204421878, + 0.0027502074372023344, + 0.00410882756114006, + 0.0025884760543704033, + 0.0033733821474015713, + 0.0036187409423291683, + 0.006603491026908159, + 0.0024462114088237286, + 0.005585170350968838, + 0.004895363934338093, + 0.010381017811596394, + 0.006995338015258312, + 0.0087182791903615, + 0.009337634779512882, + 0.007326681166887283, + 0.007256959564983845, + 0.014070069417357445, + 0.010471655987203121, + 0.021071847528219223, + 0.005681839305907488, + 0.015154367312788963, + 0.024586522951722145, + 0.00918133370578289, + 0.01513321790844202, + 0.009240723215043545, + 0.010344067588448524, + 0.020760033279657364, + 0.00516318716108799, + 0.01177515834569931, + 0.030114686116576195, + 0.006257377099245787, + 0.014551029540598392, + 0.012712073512375355, + 0.008748031221330166, + 0.11473014205694199, + 0.009293665178120136, + 0.013615249656140804, + 0.09076222032308578, + 0.005541360471397638, + 0.040592167526483536, + 0.0049003250896930695, + 0.004428206477314234, + 0.004654730204492807, + 0.003720854176208377, + 0.003970233257859945, + 0.004229401703923941, + 0.005521799903362989, + 0.005630407948046923, + 0.018462389707565308, + 0.011231918819248676, + 0.020184746012091637, + 0.01620403490960598, + 0.023443033918738365, + 0.017029372975230217, + 0.03250632807612419, + 0.012655933387577534, + 0.00366250891238451, + 0.003883359022438526, + 0.003465488087385893, + 0.0023942547850310802, + 0.002034198259934783, + 0.0034303786233067513, + 0.003012619912624359, + 0.004441999830305576, + 0.008377346210181713, + 0.004587136674672365, + 0.01828845962882042, + 0.003166403155773878, + 0.004565011244267225, + 0.011147359386086464, + 0.004078404512256384, + 0.003198573598638177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.033724840730428696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539120391011238 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012539120391011238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539120391011238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11479252576828003, + 0.06023996323347092, + 0.08770790696144104, + 0.07622817158699036, + 0.11530382186174393, + 0.11914845556020737, + 0.16440081596374512, + 0.09447453171014786, + 0.05550771206617355, + 0.11219610273838043 + ], + "credit_map": [ + 0.010137306526303291, + 0.022034931927919388, + 0.017379598692059517, + 0.012578585185110569, + 0.014650840312242508, + 0.012070877477526665, + 0.013566974550485611, + 0.01237341109663248, + 0.0055609638802707195, + 0.006134700961410999, + 0.007643782068043947, + 0.005233350675553083, + 0.006855433806777, + 0.005860099568963051, + 0.01303031761199236, + 0.009921316988766193, + 0.007525114342570305, + 0.006456849630922079, + 0.019334042444825172, + 0.012834443710744381, + 0.010367684997618198, + 0.007840079255402088, + 0.008085863664746284, + 0.015263833105564117, + 0.010299989953637123, + 0.008348972536623478, + 0.009538073092699051, + 0.006439606659114361, + 0.01368970051407814, + 0.008819867856800556, + 0.008333172649145126, + 0.010758791118860245, + 0.009185944683849812, + 0.011884928680956364, + 0.019891653209924698, + 0.008601685054600239, + 0.014632375910878181, + 0.028634654358029366, + 0.007941341027617455, + 0.014531238004565239, + 0.009853262454271317, + 0.0044370125979185104, + 0.04154668003320694, + 0.0052014486864209175, + 0.006376837380230427, + 0.03412878140807152, + 0.005505732726305723, + 0.012098696082830429, + 0.023992998525500298, + 0.015483339317142963, + 0.017839482054114342, + 0.017311427742242813, + 0.026336509734392166, + 0.01652502827346325, + 0.026134559884667397, + 0.02077748253941536, + 0.011141536757349968, + 0.006894501857459545, + 0.01466073002666235, + 0.009642401710152626, + 0.014826273545622826, + 0.010319274850189686, + 0.018322305753827095, + 0.00866751279681921, + 0.006511244922876358, + 0.005552569404244423, + 0.0075557781383395195, + 0.004828195087611675, + 0.004018763080239296, + 0.012607639655470848, + 0.005232358351349831, + 0.009201166220009327, + 0.02188740111887455, + 0.007916105911135674, + 0.0405559241771698, + 0.006188484840095043, + 0.004952194634824991, + 0.020803144201636314, + 0.004986044019460678, + 0.004906802903860807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.012539120391011238 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069907806813717 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.011069907806813717, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069907806813717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21253538131713867, + 0.08248265832662582, + 0.05636048689484596, + 0.06336696445941925, + 0.13695277273654938, + 0.14056703448295593, + 0.08919016271829605, + 0.10004245489835739, + 0.06276533752679825, + 0.05573670193552971 + ], + "credit_map": [ + 0.018711578100919724, + 0.03466755151748657, + 0.0257857758551836, + 0.021648233756422997, + 0.02299591526389122, + 0.03495942801237106, + 0.027702776715159416, + 0.02606411650776863, + 0.00847928412258625, + 0.005743362940847874, + 0.011812472715973854, + 0.007367211859673262, + 0.009573044255375862, + 0.011071269400417805, + 0.017729738727211952, + 0.010706273838877678, + 0.007449465803802013, + 0.005203831475228071, + 0.006428088992834091, + 0.007501032669097185, + 0.006915581878274679, + 0.009897246956825256, + 0.006838511675596237, + 0.006126728840172291, + 0.008486815728247166, + 0.005610449239611626, + 0.010058565065264702, + 0.0044837589375674725, + 0.006477451883256435, + 0.01315931137651205, + 0.00611785426735878, + 0.008972757495939732, + 0.009638871066272259, + 0.01197915617376566, + 0.027575653046369553, + 0.008689649403095245, + 0.013620572164654732, + 0.037573665380477905, + 0.007454039994627237, + 0.020421171560883522, + 0.012869436293840408, + 0.007935510016977787, + 0.033228550106287, + 0.007558570243418217, + 0.009761529974639416, + 0.019045475870370865, + 0.009780311025679111, + 0.04038764908909798, + 0.009646214544773102, + 0.008866602554917336, + 0.010541366413235664, + 0.009768078103661537, + 0.015993526205420494, + 0.010273012332618237, + 0.014907242730259895, + 0.009194121696054935, + 0.009828868322074413, + 0.007170976139605045, + 0.014092517085373402, + 0.01084511261433363, + 0.011966192163527012, + 0.01699931174516678, + 0.01583848148584366, + 0.013300987891852856, + 0.006516032386571169, + 0.008505113422870636, + 0.006765240337699652, + 0.006244690623134375, + 0.0057026478461921215, + 0.009090595878660679, + 0.011096117086708546, + 0.008844898082315922, + 0.007526987697929144, + 0.0097181536257267, + 0.00835735909640789, + 0.005283113103359938, + 0.007706813048571348, + 0.0055187782272696495, + 0.007751385681331158, + 0.0038741109892725945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.011069907806813717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1412133425474167, + 0.025925232097506523, + 0.09242743998765945, + 0.02943345345556736, + 0.04554005712270737, + 0.29479044675827026, + 0.05864749848842621, + 0.17133142054080963, + 0.05350746586918831, + 0.08718357235193253 + ], + "credit_map": [ + 0.011758869513869286, + 0.0232465211302042, + 0.01788513734936714, + 0.014848303981125355, + 0.016910433769226074, + 0.028287459164857864, + 0.012852409854531288, + 0.015424219891428947, + 0.0032184238079935312, + 0.0025035995058715343, + 0.0036856175865978003, + 0.0022349983919411898, + 0.003943373449146748, + 0.002764409873634577, + 0.0039588697254657745, + 0.003615941619500518, + 0.008413524366915226, + 0.0032280024606734514, + 0.034923337399959564, + 0.005597970448434353, + 0.002839235821738839, + 0.005466318689286709, + 0.0037462895270437002, + 0.028212768957018852, + 0.004106797277927399, + 0.002794734202325344, + 0.003614333923906088, + 0.003004851983860135, + 0.003708295989781618, + 0.004347972571849823, + 0.0031928839161992073, + 0.00466358894482255, + 0.004850146826356649, + 0.003637595335021615, + 0.00964483805000782, + 0.0023524716962128878, + 0.004757882561534643, + 0.010742559097707272, + 0.0020435538608580828, + 0.007511013187468052, + 0.009416284039616585, + 0.007859045639634132, + 0.09779298305511475, + 0.007373396772891283, + 0.011538955383002758, + 0.12334883958101273, + 0.003923737443983555, + 0.03353722766041756, + 0.0074932994320988655, + 0.004830228164792061, + 0.006837319582700729, + 0.0060272784903645515, + 0.011058746837079525, + 0.00577474245801568, + 0.009216446429491043, + 0.007409445475786924, + 0.019668899476528168, + 0.013090120628476143, + 0.02175663597881794, + 0.01889912597835064, + 0.026663878932595253, + 0.018681280314922333, + 0.03703581914305687, + 0.015535683371126652, + 0.006110926624387503, + 0.004953587427735329, + 0.008475182577967644, + 0.004655982367694378, + 0.003998030908405781, + 0.009197888895869255, + 0.007337252143770456, + 0.008778625167906284, + 0.009513594210147858, + 0.011289369314908981, + 0.021628964692354202, + 0.009930788539350033, + 0.0052117411978542805, + 0.01668502762913704, + 0.006148169282823801, + 0.006775916554033756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.004635422024875879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, fixed wide-angle shot, shot from above, captures the moment when the swirling, inky clouds of the heavens are suddenly ripped apart by an invisible force. The Monkey King, adorned with an intricately carved, majestically shimmering phoenix-wing purple-gold crown, draped in golden chain mail armor, and tied with an ochre-yellow tiger-skin skirt, bursts explosively from beneath the clouds. His face is defiant, his eyes wide open, radiating a piercing golden light. His bared teeth utter a silent war cry. He clutches the radiant golden Ruyi Jingu Bang in both hands and swings it forward with thunderous force. The energy halo surrounding the club instantly pierces the thick wall of cloud, sending countless shattered clouds into a churning wave that scatters in all directions, forming a vast, ever-expanding vortex-like void behind him. The whole picture is full of epic power and the magnificence of breaking free from constraints. The golden sacred light forms a sharp contrast with the dark clouds, creating a mythical, grand and fierce battle atmosphere.", + "seed": 3660, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_5595_seed_3660_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": -0.1885360856540501, + "vlm_logprob": -0.1885360856540501, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.1885360856540501, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9010749459266663 + ], + "opposite_logprob_tokens": [ + -0.5260749459266663 + ], + "logprob_reward": -0.9010749459266663, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9010749459266663, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16842715442180634, + 0.050950340926647186, + 0.03607563674449921, + 0.01302967220544815, + 0.06535713374614716, + 0.03123190999031067, + 0.03255711868405342, + 0.5215193033218384, + 0.041271451860666275, + 0.03958022966980934 + ], + "credit_map": [ + 0.018753720447421074, + 0.03071708045899868, + 0.04363119229674339, + 0.01979510858654976, + 0.014431286603212357, + 0.015661559998989105, + 0.014252287335693836, + 0.011184915900230408, + 0.003761348547413945, + 0.0052689299918711185, + 0.006153432186692953, + 0.00736142648383975, + 0.008045781403779984, + 0.005837167147547007, + 0.009821698069572449, + 0.004700557328760624, + 0.003840396646410227, + 0.003813687711954117, + 0.005691046826541424, + 0.00332582532428205, + 0.004404051695019007, + 0.005945589393377304, + 0.003595107700675726, + 0.005459929816424847, + 0.001801392063498497, + 0.001227133208885789, + 0.0021980691235512495, + 0.0010245025623589754, + 0.0013704600278288126, + 0.0027887371834367514, + 0.0012691409792751074, + 0.0013502370566129684, + 0.00610645255073905, + 0.00252806325443089, + 0.019383922219276428, + 0.0028707722667604685, + 0.004742656834423542, + 0.024357883259654045, + 0.0029083879198879004, + 0.002459001960232854, + 0.004136078059673309, + 0.002554381964728236, + 0.0034231285098940134, + 0.0023404087405651808, + 0.0033989730291068554, + 0.009147663600742817, + 0.0024538489524275064, + 0.003777427366003394, + 0.005471523851156235, + 0.0023191063664853573, + 0.0064254095777869225, + 0.002294727601110935, + 0.0034855131525546312, + 0.004232665058225393, + 0.002912436379119754, + 0.0054157390259206295, + 0.1159287691116333, + 0.02347639761865139, + 0.04894481226801872, + 0.01183465588837862, + 0.03643332049250603, + 0.2180652618408203, + 0.011763690039515495, + 0.05507243424654007, + 0.0039142947643995285, + 0.004948627669364214, + 0.0051365382969379425, + 0.006188392173498869, + 0.004430387634783983, + 0.006800714880228043, + 0.004623758140951395, + 0.005228737369179726, + 0.00482933921739459, + 0.004661290906369686, + 0.005965367890894413, + 0.0038215508684515953, + 0.007687967736274004, + 0.004516477230936289, + 0.004046823363751173, + 0.004051412455737591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.9010749459266663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0892360731959343, + 0.07357639074325562, + 0.3103463649749756, + 0.12289511412382126, + 0.1578834354877472, + 0.03741487115621567, + 0.04624026268720627, + 0.07344509661197662, + 0.0415373370051384, + 0.04742508381605148 + ], + "credit_map": [ + 0.011762796901166439, + 0.010898148640990257, + 0.01639959216117859, + 0.012591274455189705, + 0.008426299318671227, + 0.01142830215394497, + 0.009606375358998775, + 0.008123265579342842, + 0.006686138454824686, + 0.00914761982858181, + 0.007801603525876999, + 0.011764921247959137, + 0.013209193013608456, + 0.008583779446780682, + 0.01185902114957571, + 0.004524113144725561, + 0.03045644611120224, + 0.028131889179348946, + 0.051125410944223404, + 0.02219937928020954, + 0.03880929946899414, + 0.05472726747393608, + 0.026463089510798454, + 0.05843353644013405, + 0.009918604977428913, + 0.011537587270140648, + 0.018384620547294617, + 0.008945746347308159, + 0.02596495859324932, + 0.01862492598593235, + 0.011901406571269035, + 0.017617242410779, + 0.013995390385389328, + 0.009417465887963772, + 0.03414643183350563, + 0.009571726433932781, + 0.021401893347501755, + 0.0421922504901886, + 0.017353815957903862, + 0.009804442524909973, + 0.0054721347987651825, + 0.002781229792162776, + 0.004723503254354, + 0.002868901938199997, + 0.00437315134331584, + 0.010966554284095764, + 0.0024556308053433895, + 0.0037737598177045584, + 0.004587326664477587, + 0.00571416737511754, + 0.006500972900539637, + 0.004645445384085178, + 0.008064434863626957, + 0.005785771645605564, + 0.005141464993357658, + 0.005800671875476837, + 0.01600826531648636, + 0.0054082684218883514, + 0.007639577146619558, + 0.00467795180156827, + 0.005103779025375843, + 0.02396243065595627, + 0.003246073145419359, + 0.007398746442049742, + 0.0037419151049107313, + 0.0045412201434373856, + 0.006227363832294941, + 0.007654756773263216, + 0.003846073057502508, + 0.006261569447815418, + 0.00470346212387085, + 0.0045609744265675545, + 0.004885215312242508, + 0.006523525342345238, + 0.005388185381889343, + 0.005877717863768339, + 0.008241130970418453, + 0.005248561501502991, + 0.007079584989696741, + 0.00418115546926856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.014206582680344582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076981201767921 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076981201767921, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076981201767921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043554119765758514, + 0.04731050506234169, + 0.1216803640127182, + 0.04415331035852432, + 0.06409580260515213, + 0.0778367668390274, + 0.03606279566884041, + 0.49997079372406006, + 0.029980739578604698, + 0.03535480052232742 + ], + "credit_map": [ + 0.004618115723133087, + 0.004787022713571787, + 0.004194046836346388, + 0.010021964088082314, + 0.00591895030811429, + 0.005748421419411898, + 0.004921712446957827, + 0.0033438855316489935, + 0.004259984474629164, + 0.0056160525418818, + 0.0052413614466786385, + 0.006651055999100208, + 0.007275847252458334, + 0.0065933093428611755, + 0.007063033990561962, + 0.004609859548509121, + 0.015062445774674416, + 0.010256584733724594, + 0.02001868560910225, + 0.008632022887468338, + 0.012372319586575031, + 0.02255498431622982, + 0.011251355521380901, + 0.021531973034143448, + 0.004626968875527382, + 0.0064322990365326405, + 0.005885046906769276, + 0.0042747147381305695, + 0.009395691566169262, + 0.0033337504137307405, + 0.0059566074050962925, + 0.004248226527124643, + 0.004961686208844185, + 0.004033959936350584, + 0.01727760210633278, + 0.0037638447247445583, + 0.006120807025581598, + 0.020645558834075928, + 0.004307843744754791, + 0.002984496997669339, + 0.010456254705786705, + 0.009186691604554653, + 0.00867282785475254, + 0.004865922033786774, + 0.010610087774693966, + 0.018560755997896194, + 0.0072110979817807674, + 0.008273130282759666, + 0.005683776922523975, + 0.0037646552082151175, + 0.005210567731410265, + 0.003430863842368126, + 0.00402672216296196, + 0.004690816625952721, + 0.0043366290628910065, + 0.004918764811009169, + 0.10305499285459518, + 0.02208676189184189, + 0.046457212418317795, + 0.009181512519717216, + 0.03680385649204254, + 0.21172621846199036, + 0.00988447293639183, + 0.0607757605612278, + 0.0029061094392091036, + 0.0033443698193877935, + 0.0039965989999473095, + 0.004927457310259342, + 0.004422520752996206, + 0.003540267003700137, + 0.0031549776904284954, + 0.003688436234369874, + 0.0037188739515841007, + 0.005169362295418978, + 0.004738610703498125, + 0.004005291033536196, + 0.005381960887461901, + 0.0043159471824765205, + 0.004547241143882275, + 0.003477511927485466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.011076981201767921 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070851236581802 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070851236581802, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070851236581802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05077644810080528, + 0.03274804726243019, + 0.04518680274486542, + 0.07570909708738327, + 0.05697586014866829, + 0.03025989979505539, + 0.021120933815836906, + 0.6490887999534607, + 0.01861615851521492, + 0.019517960026860237 + ], + "credit_map": [ + 0.005651495885103941, + 0.008599109947681427, + 0.011947320774197578, + 0.008401009254157543, + 0.004390081390738487, + 0.004222959280014038, + 0.004669967107474804, + 0.0028945053927600384, + 0.002994273789227009, + 0.003742281114682555, + 0.0043187132105231285, + 0.005778574384748936, + 0.004630813375115395, + 0.0041657923720777035, + 0.0048432862386107445, + 0.0022743127774447203, + 0.00609068525955081, + 0.0059263682924211025, + 0.005961742252111435, + 0.0038762823678553104, + 0.006450934335589409, + 0.00621009711176157, + 0.004648905247449875, + 0.006021787878125906, + 0.009830964729189873, + 0.00853797234594822, + 0.013408522121608257, + 0.0061463890597224236, + 0.0056026894599199295, + 0.011284020729362965, + 0.007511822506785393, + 0.013386721722781658, + 0.005149118602275848, + 0.005220668390393257, + 0.01346119586378336, + 0.0030796716455370188, + 0.004924253094941378, + 0.018407750874757767, + 0.0039000229444354773, + 0.002833176637068391, + 0.003944810014218092, + 0.003750215983018279, + 0.003909520339220762, + 0.002688511274755001, + 0.0035898257046937943, + 0.00645084772258997, + 0.002778387861326337, + 0.003147780429571867, + 0.0022262735292315483, + 0.002098222030326724, + 0.003172728233039379, + 0.0023061048705130816, + 0.0023894018959254026, + 0.002963453996926546, + 0.0029820415657013655, + 0.002982707694172859, + 0.1339370310306549, + 0.02944045513868332, + 0.060350220650434494, + 0.012137503363192081, + 0.0473875030875206, + 0.27606695890426636, + 0.012510225176811218, + 0.07725891470909119, + 0.0017556124366819859, + 0.002166602062061429, + 0.0026092580519616604, + 0.002630874514579773, + 0.002074867719784379, + 0.0030935609247535467, + 0.00200972193852067, + 0.002275659702718258, + 0.0020463517867028713, + 0.0025324251037091017, + 0.002738090232014656, + 0.0016714924713596702, + 0.003631039522588253, + 0.002239424968138337, + 0.002323842840269208, + 0.0023352953139692545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.011070851236581802 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12386906892061234, + 0.0768357589840889, + 0.035524170845746994, + 0.014320513233542442, + 0.11906979233026505, + 0.02218104526400566, + 0.03293577581644058, + 0.48262184858322144, + 0.04340631887316704, + 0.04923567920923233 + ], + "credit_map": [ + 0.014161881059408188, + 0.015062933787703514, + 0.02246692404150963, + 0.021900741383433342, + 0.010680040344595909, + 0.015142185613512993, + 0.013357561081647873, + 0.01109679788351059, + 0.005927673075348139, + 0.008425780571997166, + 0.008640473708510399, + 0.008993957191705704, + 0.012236305512487888, + 0.01207813061773777, + 0.013781480491161346, + 0.006751959677785635, + 0.0034980603959411383, + 0.0033670251723378897, + 0.0055149514228105545, + 0.0035607172176241875, + 0.005126967094838619, + 0.005567758344113827, + 0.004026817157864571, + 0.004861874505877495, + 0.0016967186238616705, + 0.0017533849459141493, + 0.0017766152741387486, + 0.0012084586778655648, + 0.0019577606581151485, + 0.002661662409082055, + 0.0018472918309271336, + 0.001418621395714581, + 0.011043468490242958, + 0.006397459656000137, + 0.029920732602477074, + 0.006516453810036182, + 0.012294449843466282, + 0.03919692710042, + 0.008089141920208931, + 0.005611163564026356, + 0.003165267640724778, + 0.0021808596793562174, + 0.0026067078579217196, + 0.0019669439643621445, + 0.002220995957031846, + 0.0051227654330432415, + 0.0020732295233756304, + 0.002844274742528796, + 0.00513428682461381, + 0.0030888793990015984, + 0.006626591086387634, + 0.0023895162157714367, + 0.0033653986174613237, + 0.004693937953561544, + 0.002559622749686241, + 0.005077544134110212, + 0.10303650796413422, + 0.024776531383395195, + 0.0442824587225914, + 0.012972327880561352, + 0.03361076861619949, + 0.1999569535255432, + 0.010420802049338818, + 0.05356549099087715, + 0.004295418504625559, + 0.005138070788234472, + 0.006048102863132954, + 0.0075626615434885025, + 0.00555697176605463, + 0.005680450238287449, + 0.004245595540851355, + 0.004879046231508255, + 0.005861107259988785, + 0.006305747199803591, + 0.006816528737545013, + 0.0054770871065557, + 0.0075676944106817245, + 0.005350620020180941, + 0.006736284587532282, + 0.005120614543557167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.005251067224889994 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low, fixed wide-angle shot, shot from above, captures the moment when the swirling, inky clouds of the heavens are suddenly ripped apart by an invisible force. The Monkey King, adorned with an intricately carved, majestically shimmering phoenix-wing purple-gold crown, draped in golden chain mail armor, and tied with an ochre-yellow tiger-skin skirt, bursts explosively from beneath the clouds. His face is defiant, his eyes wide open, radiating a piercing golden light. His bared teeth utter a silent war cry. He clutches the radiant golden Ruyi Jingu Bang in both hands and swings it forward with thunderous force. The energy halo surrounding the club instantly pierces the thick wall of cloud, sending countless shattered clouds into a churning wave that scatters in all directions, forming a vast, ever-expanding vortex-like void behind him. The whole picture is full of epic power and the magnificence of breaking free from constraints. The golden sacred light forms a sharp contrast with the dark clouds, creating a mythical, grand and fierce battle atmosphere.", + "seed": 3661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_6/index_5595_seed_3661_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": -0.13890368063002825, + "vlm_logprob": -0.13890368063002825, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.13890368063002825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.641127347946167 + ], + "opposite_logprob_tokens": [ + -0.766127347946167 + ], + "logprob_reward": -0.641127347946167, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.641127347946167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3213307559490204, + 0.13782304525375366, + 0.03351127356290817, + 0.026927238330245018, + 0.028304213657975197, + 0.035683173686265945, + 0.1980716735124588, + 0.11076831817626953, + 0.05574628710746765, + 0.05183401331305504 + ], + "credit_map": [ + 0.03585805743932724, + 0.03208192065358162, + 0.036600593477487564, + 0.07407122850418091, + 0.04252190887928009, + 0.037566863000392914, + 0.03225667402148247, + 0.030373532325029373, + 0.01368593331426382, + 0.012783733196556568, + 0.014544869773089886, + 0.013593928888440132, + 0.025529874488711357, + 0.02798755280673504, + 0.015542478300631046, + 0.014154680073261261, + 0.0034293814096599817, + 0.003124167677015066, + 0.005531940143555403, + 0.002959209494292736, + 0.004801678471267223, + 0.005022880621254444, + 0.0031995121389627457, + 0.005442505236715078, + 0.003148141782730818, + 0.0024808815214782953, + 0.004384438507258892, + 0.0030296831391751766, + 0.002675836905837059, + 0.0037395181134343147, + 0.004192059859633446, + 0.0032766791991889477, + 0.0027717293705791235, + 0.002960313344374299, + 0.003684195689857006, + 0.0031889949459582567, + 0.00262085092253983, + 0.004319996573030949, + 0.002395063638687134, + 0.0063630701042711735, + 0.005011726636439562, + 0.002953415736556053, + 0.004942853935062885, + 0.005210919305682182, + 0.004466813988983631, + 0.004287056624889374, + 0.0039335149340331554, + 0.004876875318586826, + 0.013036967255175114, + 0.008726604282855988, + 0.044441111385822296, + 0.004674033727496862, + 0.00771714560687542, + 0.0405564084649086, + 0.006449598353356123, + 0.0724697932600975, + 0.035607434809207916, + 0.005856890231370926, + 0.040718644857406616, + 0.004925799556076527, + 0.004583616741001606, + 0.0067895883694291115, + 0.004707226529717445, + 0.0075791338458657265, + 0.005128397606313229, + 0.006821180693805218, + 0.007933479733765125, + 0.010968372225761414, + 0.004476397763937712, + 0.007217267528176308, + 0.005148416385054588, + 0.008052772842347622, + 0.003673611907288432, + 0.011095991358160973, + 0.0047110700979828835, + 0.0077034845016896725, + 0.005004577804356813, + 0.0033034600783139467, + 0.011363890022039413, + 0.004977930337190628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.641127347946167 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033720921725034714 + ], + "opposite_logprob_tokens": [ + -3.4087209701538086 + ], + "logprob_reward": -0.033720921725034714, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033720921725034714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0457313247025013, + 0.018806463107466698, + 0.012873164378106594, + 0.024430910125374794, + 0.012257971800863743, + 0.030993768945336342, + 0.18432192504405975, + 0.6460226774215698, + 0.013676458969712257, + 0.010885348543524742 + ], + "credit_map": [ + 0.005109363701194525, + 0.004771535284817219, + 0.0038784455973654985, + 0.005958783905953169, + 0.009366732090711594, + 0.0065879542380571365, + 0.004876357037574053, + 0.005182153545320034, + 0.0020299262832850218, + 0.0016041775234043598, + 0.0017564949812367558, + 0.0017267426010221243, + 0.0031898366287350655, + 0.004245900083333254, + 0.0017237999709323049, + 0.0025295850355178118, + 0.0010054270969703794, + 0.0013768647331744432, + 0.0018276141490787268, + 0.001325099146924913, + 0.0018139858730137348, + 0.0016954205930233002, + 0.0014706563670188189, + 0.002358095720410347, + 0.002385843778029084, + 0.002818088745698333, + 0.0036903175059705973, + 0.004043269902467728, + 0.0027388378512114286, + 0.0029313983395695686, + 0.0028297032695263624, + 0.0029934500344097614, + 0.0012508431682363153, + 0.0012819309486076236, + 0.0016796586569398642, + 0.0017892206087708473, + 0.001450949115678668, + 0.0015402695862576365, + 0.0011689767707139254, + 0.002096123294904828, + 0.004014972597360611, + 0.002835998311638832, + 0.0032696255948394537, + 0.004366959910839796, + 0.003873404348269105, + 0.004200742579996586, + 0.0031267469748854637, + 0.005305319558829069, + 0.01739547774195671, + 0.01630573347210884, + 0.03744906559586525, + 0.0068366448394954205, + 0.010274680331349373, + 0.04513584449887276, + 0.0057037994265556335, + 0.04522067308425903, + 0.2513620853424072, + 0.036013707518577576, + 0.2630484700202942, + 0.02455432154238224, + 0.010768883861601353, + 0.020084688439965248, + 0.02240797132253647, + 0.01778254099190235, + 0.0015473929233849049, + 0.001835431670770049, + 0.0024207786191254854, + 0.002070491900667548, + 0.0011270307004451752, + 0.0010582657996565104, + 0.0013509112177416682, + 0.002266156952828169, + 0.0012481738813221455, + 0.002272901823744178, + 0.0012285972479730844, + 0.001632815576158464, + 0.0011335466988384724, + 0.0008650000672787428, + 0.00161732395645231, + 0.0008869890589267015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.033720921725034714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.037710271775722504, + 0.01634828932583332, + 0.014261365868151188, + 0.03105483390390873, + 0.011963513679802418, + 0.03351908549666405, + 0.20696188509464264, + 0.6194524765014648, + 0.008279423229396343, + 0.020448880270123482 + ], + "credit_map": [ + 0.00208050524815917, + 0.0036008108872920275, + 0.0030091472435742617, + 0.004732625558972359, + 0.01204682793468237, + 0.006839882116764784, + 0.0031538126058876514, + 0.002246660878881812, + 0.0017412262968719006, + 0.0015484001487493515, + 0.0018000229028984904, + 0.0025563761591911316, + 0.0027893544174730778, + 0.0021207635290920734, + 0.001978972228243947, + 0.0018131734104827046, + 0.0016110549913719296, + 0.0012227102415636182, + 0.002337734680622816, + 0.0014820977812632918, + 0.0025104552041739225, + 0.0017491086618974805, + 0.001178582664579153, + 0.0021696218755096197, + 0.004531285259872675, + 0.0039569418877363205, + 0.0037686272989958525, + 0.003705279203131795, + 0.0034400722943246365, + 0.003664053278043866, + 0.004234110005199909, + 0.003754465375095606, + 0.001667920034378767, + 0.0011654238915070891, + 0.0014410329749807715, + 0.0017320034094154835, + 0.0017262225737795234, + 0.0012944003101438284, + 0.0012189011322334409, + 0.001717609353363514, + 0.003927723970264196, + 0.003456464735791087, + 0.004619398154318333, + 0.004529302939772606, + 0.005752762779593468, + 0.002938011894002557, + 0.0044202785938978195, + 0.0038751447573304176, + 0.026648659259080887, + 0.024547459557652473, + 0.07642698287963867, + 0.008910110220313072, + 0.014554601162672043, + 0.0357176698744297, + 0.005373162217438221, + 0.014783252961933613, + 0.24538785219192505, + 0.03553315997123718, + 0.2512360215187073, + 0.023235218599438667, + 0.009838299825787544, + 0.016470523551106453, + 0.02072162553668022, + 0.017029758542776108, + 0.0009529082453809679, + 0.0013755219988524914, + 0.0011368001578375697, + 0.001227215281687677, + 0.0009309276356361806, + 0.0006671976298093796, + 0.0007225798326544464, + 0.0012662730878219008, + 0.0014417078346014023, + 0.004762931261211634, + 0.0017727199010550976, + 0.002813762752339244, + 0.0020194100216031075, + 0.00145920948125422, + 0.003924435470253229, + 0.0022547044791281223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.00978483073413372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634472541511059 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634472541511059, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634472541511059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05206770822405815, + 0.04981851950287819, + 0.036258257925510406, + 0.07768838107585907, + 0.04110175743699074, + 0.11106829345226288, + 0.26189595460891724, + 0.3051026165485382, + 0.03985997661948204, + 0.025138583034276962 + ], + "credit_map": [ + 0.004468832165002823, + 0.0060075619257986546, + 0.0057846796698868275, + 0.009961164556443691, + 0.007745753973722458, + 0.006769267376512289, + 0.004853061866015196, + 0.006477386225014925, + 0.004271051846444607, + 0.004530795384198427, + 0.00476757762953639, + 0.005612053908407688, + 0.011102331802248955, + 0.008605677634477615, + 0.006591045297682285, + 0.0043379878625273705, + 0.0039477357640862465, + 0.004189628176391125, + 0.005150387529283762, + 0.0039022399578243494, + 0.006859151180833578, + 0.004313242621719837, + 0.0040558562614023685, + 0.0038400180637836456, + 0.008100931532680988, + 0.009290214627981186, + 0.01326771266758442, + 0.014115416444838047, + 0.008542938157916069, + 0.008389618247747421, + 0.008235123939812183, + 0.007746424991637468, + 0.007515767589211464, + 0.003608213970437646, + 0.006425416097044945, + 0.005020510405302048, + 0.004391992464661598, + 0.004503477364778519, + 0.0036928195040673018, + 0.005943558178842068, + 0.01961682178080082, + 0.008468827232718468, + 0.0142753841355443, + 0.014456700533628464, + 0.011241891421377659, + 0.013669434934854507, + 0.012923751026391983, + 0.016415487974882126, + 0.016915427520871162, + 0.015810415148735046, + 0.01515475194901228, + 0.0076994094997644424, + 0.008646887727081776, + 0.07952390611171722, + 0.01062871515750885, + 0.10751643031835556, + 0.11073429137468338, + 0.016173116862773895, + 0.1221645399928093, + 0.011949844658374786, + 0.005914965178817511, + 0.014700562693178654, + 0.011700905859470367, + 0.01176439132541418, + 0.004098717588931322, + 0.0069699413143098354, + 0.006138259079307318, + 0.005301793105900288, + 0.0046098302118480206, + 0.0033606078941375017, + 0.0037666172720491886, + 0.0056142122484743595, + 0.002261318266391754, + 0.00474313460290432, + 0.003042828058823943, + 0.0034650193993002176, + 0.0025058197788894176, + 0.0024408558383584023, + 0.003934625536203384, + 0.002744981786236167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.004634472541511059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low, fixed wide-angle view captured from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear an intricately carved purple-gold phoenix-wing crown and golden chain mail armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's face defiant with wide eyes radiating piercing golden light and bared teeth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character clutch a radiant golden Ruyi Jingu Bang and swing it forward with force?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the action show the character bursting from dark clouds to create a vortex-like void of shattered clouds behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14034798741340637, + 0.0448528490960598, + 0.028973357751965523, + 0.033760689198970795, + 0.014980307780206203, + 0.022669585421681404, + 0.219722181558609, + 0.26511675119400024, + 0.1015944704413414, + 0.12798182666301727 + ], + "credit_map": [ + 0.013395534828305244, + 0.015302227810025215, + 0.01582307368516922, + 0.024973250925540924, + 0.023810341954231262, + 0.019260115921497345, + 0.012635213322937489, + 0.01514823455363512, + 0.005033391993492842, + 0.004516240209341049, + 0.004593938123434782, + 0.004115506075322628, + 0.008064229972660542, + 0.008455880917608738, + 0.005203010514378548, + 0.00487064803019166, + 0.00287242466583848, + 0.003231485839933157, + 0.004125608131289482, + 0.0029062086250633, + 0.005024767015129328, + 0.003658822039142251, + 0.0037295513320714235, + 0.0034244901034981012, + 0.0037617019843310118, + 0.0028710586484521627, + 0.007781185209751129, + 0.0025629922747612, + 0.0031647353898733854, + 0.004864371381700039, + 0.0022512476425617933, + 0.006503397133201361, + 0.0023581143468618393, + 0.0017378374468535185, + 0.001675424282439053, + 0.0017336303135380149, + 0.0021070523653179407, + 0.0016495226882398129, + 0.0014712840784341097, + 0.0022474422585219145, + 0.003997109364718199, + 0.0021571258548647165, + 0.0036122698802500963, + 0.00286012701690197, + 0.002434327732771635, + 0.002242607530206442, + 0.002076424891129136, + 0.003289592918008566, + 0.012027760036289692, + 0.010589043609797955, + 0.01783524826169014, + 0.005116517189890146, + 0.006435961928218603, + 0.06515197455883026, + 0.008420680649578571, + 0.09414499998092651, + 0.07273947447538376, + 0.013927279971539974, + 0.1005532294511795, + 0.013253424316644669, + 0.007184659130871296, + 0.026761915534734726, + 0.012146665714681149, + 0.018550122156739235, + 0.008943559601902962, + 0.01881531998515129, + 0.013484399765729904, + 0.02006886899471283, + 0.0075786178931593895, + 0.011878158897161484, + 0.009801913984119892, + 0.011023632250726223, + 0.009949429892003536, + 0.02813367173075676, + 0.013685633428394794, + 0.019689463078975677, + 0.01593221351504326, + 0.007068750448524952, + 0.022842759266495705, + 0.01067990530282259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.005250830203294754 + } + ], + "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/0000110/group_7/index_tensor([1246, 1246])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_tensor([1246, 1246])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6dcc2fac38a32b58f954eba89d48967b04115f12 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_tensor([1246, 1246])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, focusing on an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in the turbulent, murky, yellow-brown water. Soft, diffuse morning light penetrates the mist, illuminating his sparse, short, graying hair and equally graying beard. Drops of water trickle down his furrowed brow and cheeks strained with effort. Bare to the waist, his muscles, hardened by years of toil, now appear limp. He leans back, struggling to keep his feet steady in the muddy water beneath the surface, bracing himself against the powerful current. He clutched a rough, frayed hemp rope tightly in his hands. With every tremor of his muscles and every heavy breath, he dragged a mysterious old wooden box, partially covered in weeds and silt, inch by inch from the surging river. The box's edges, inlaid with rusted brass corners, gleamed dimly in the dim light, while icy water flowed continuously from the cracks in the box. The entire scene was filled with a sense of primal power, tense suspense, and the mysterious atmosphere of discovering the unknown. The overall color tone was cold, with a primitive and solemn texture.", + "seed": 5329, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5329_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": -0.11930485414341092, + "vlm_logprob": -0.11930485414341092, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.11930485414341092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.15603111684322357 + ], + "opposite_logprob_tokens": [ + -2.5310311317443848 + ], + "logprob_reward": -0.15603111684322357, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.15603111684322357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16716308891773224, + 0.08458944410085678, + 0.08156227320432663, + 0.15623222291469574, + 0.060023773461580276, + 0.07516689598560333, + 0.0994594618678093, + 0.06275998800992966, + 0.13010412454605103, + 0.08293874561786652 + ], + "credit_map": [ + 0.04202495142817497, + 0.020927276462316513, + 0.014184942469000816, + 0.010001561604440212, + 0.017903124913573265, + 0.027646446600556374, + 0.018627552315592766, + 0.015847232192754745, + 0.020400989800691605, + 0.011256716214120388, + 0.008857290260493755, + 0.0053375218994915485, + 0.007565704640001059, + 0.013701974414288998, + 0.008208724670112133, + 0.009260524995625019, + 0.019499002024531364, + 0.00922198873013258, + 0.00898903887718916, + 0.003495217999443412, + 0.008953922428190708, + 0.012970038689672947, + 0.007390693761408329, + 0.011042368598282337, + 0.04792466759681702, + 0.014798651449382305, + 0.010476679541170597, + 0.007103262934833765, + 0.011782052926719189, + 0.04840763658285141, + 0.0065620834939181805, + 0.009177184663712978, + 0.016641784459352493, + 0.006407710257917643, + 0.006421499885618687, + 0.004600923508405685, + 0.004017885774374008, + 0.012030260637402534, + 0.004252102226018906, + 0.0056516011245548725, + 0.008983073756098747, + 0.006745908409357071, + 0.013337434269487858, + 0.008195985108613968, + 0.007688488345593214, + 0.007956728339195251, + 0.015273049473762512, + 0.006986228749155998, + 0.010035085491836071, + 0.009397047571837902, + 0.02495601959526539, + 0.009731941856443882, + 0.012643016874790192, + 0.011603436432778835, + 0.01048693060874939, + 0.010605989024043083, + 0.006540162023156881, + 0.007955568842589855, + 0.0075615448877215385, + 0.010406692512333393, + 0.007124427706003189, + 0.008113562129437923, + 0.009792509488761425, + 0.005265521816909313, + 0.012598331086337566, + 0.02030766010284424, + 0.01922191120684147, + 0.01917906291782856, + 0.014386248774826527, + 0.015198932029306889, + 0.015761621296405792, + 0.013450352475047112, + 0.012507838197052479, + 0.007210391573607922, + 0.010493340902030468, + 0.007482881657779217, + 0.008406031876802444, + 0.01745576784014702, + 0.010189979337155819, + 0.009192517958581448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.15603111684322357 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.3486550748348236 + ], + "opposite_logprob_tokens": [ + -1.223655104637146 + ], + "logprob_reward": -0.3486550748348236, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486550748348236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17417021095752716, + 0.057207122445106506, + 0.11807034909725189, + 0.07953732460737228, + 0.06783440709114075, + 0.08150060474872589, + 0.06725393235683441, + 0.09111065417528152, + 0.1306517869234085, + 0.13266359269618988 + ], + "credit_map": [ + 0.04965883865952492, + 0.021614503115415573, + 0.019311098381876945, + 0.008432864211499691, + 0.014591802842915058, + 0.030151376500725746, + 0.013827632181346416, + 0.016582094132900238, + 0.013634455390274525, + 0.007062381133437157, + 0.006587734911590815, + 0.0038687451742589474, + 0.006220092996954918, + 0.008864258415997028, + 0.005438417196273804, + 0.0055310409516096115, + 0.02535947971045971, + 0.013996579684317112, + 0.011461518704891205, + 0.006826573051512241, + 0.018862776458263397, + 0.01747608184814453, + 0.013103282079100609, + 0.010984061285853386, + 0.014371947385370731, + 0.0092953285202384, + 0.010659773834049702, + 0.00511175487190485, + 0.008881092071533203, + 0.013417697511613369, + 0.005218919366598129, + 0.012580819427967072, + 0.01678040623664856, + 0.007410531397908926, + 0.006534578278660774, + 0.0040983157232403755, + 0.006975255906581879, + 0.015518778935074806, + 0.004751404281705618, + 0.005765137262642384, + 0.014288618229329586, + 0.009392171166837215, + 0.008439426310360432, + 0.007852163165807724, + 0.014182116836309433, + 0.010205051861703396, + 0.00976728368550539, + 0.007373773027211428, + 0.008727004751563072, + 0.008656060323119164, + 0.014896309934556484, + 0.008501890115439892, + 0.0075208041816949844, + 0.00766231631860137, + 0.005632832180708647, + 0.00565672479569912, + 0.010205640457570553, + 0.010075115598738194, + 0.013879257254302502, + 0.016378507018089294, + 0.010293579660356045, + 0.011716837994754314, + 0.011366162449121475, + 0.007195562124252319, + 0.014535505324602127, + 0.02239060588181019, + 0.02015296369791031, + 0.018758876249194145, + 0.013260599225759506, + 0.0155789190903306, + 0.014804229140281677, + 0.011170094832777977, + 0.01715748757123947, + 0.015492510050535202, + 0.016300659626722336, + 0.01610899530351162, + 0.01113073155283928, + 0.018155811354517937, + 0.02100609429180622, + 0.017311308532953262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.3486550748348236 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.01819690316915512 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.01819690316915512, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819690316915512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23036402463912964, + 0.06140250712633133, + 0.14543195068836212, + 0.19238176941871643, + 0.08970479667186737, + 0.06230591982603073, + 0.053373225033283234, + 0.055631402879953384, + 0.06463539600372314, + 0.044768963009119034 + ], + "credit_map": [ + 0.05376669391989708, + 0.03704490885138512, + 0.02736222930252552, + 0.014032379724085331, + 0.02248629555106163, + 0.03725099936127663, + 0.016260379925370216, + 0.022160138934850693, + 0.011559267528355122, + 0.01012883335351944, + 0.0060685742646455765, + 0.00458309194073081, + 0.008737776428461075, + 0.009160256013274193, + 0.006266747135668993, + 0.004897961858659983, + 0.028076887130737305, + 0.01741841621696949, + 0.017001576721668243, + 0.01120106689631939, + 0.01652107574045658, + 0.01877739652991295, + 0.014662967063486576, + 0.021772567182779312, + 0.05042535439133644, + 0.021927248686552048, + 0.030072322115302086, + 0.01416429691016674, + 0.015624402090907097, + 0.040447499603033066, + 0.01049881987273693, + 0.009221828542649746, + 0.026635002344846725, + 0.008616643957793713, + 0.009877650998532772, + 0.00662336265668273, + 0.007151117082685232, + 0.01645539328455925, + 0.00695182615891099, + 0.007393799722194672, + 0.007171666715294123, + 0.00853804498910904, + 0.01094374805688858, + 0.006598701234906912, + 0.010136701166629791, + 0.007484589237719774, + 0.006728903856128454, + 0.0047035678289830685, + 0.0042232028208673, + 0.005012724082916975, + 0.005991205107420683, + 0.006860766559839249, + 0.009780246764421463, + 0.011644254438579082, + 0.005149439908564091, + 0.0047113848850131035, + 0.009488734416663647, + 0.004809068515896797, + 0.007004015147686005, + 0.008147208951413631, + 0.006651378236711025, + 0.007477881852537394, + 0.006945644039660692, + 0.005107474513351917, + 0.006370199844241142, + 0.00873818900436163, + 0.010732077062129974, + 0.007803237531334162, + 0.007100710645318031, + 0.006516935769468546, + 0.010907985270023346, + 0.00646606320515275, + 0.006329397205263376, + 0.004461641423404217, + 0.004450703971087933, + 0.0043749515898525715, + 0.0038314631674438715, + 0.006449337583035231, + 0.005945482291281223, + 0.00892598181962967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.01819690316915512 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.0036183618940413 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036183618940413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036183618940413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.244656041264534, + 0.044677816331386566, + 0.050555504858493805, + 0.1102217584848404, + 0.08519896864891052, + 0.038387276232242584, + 0.04055528715252876, + 0.0429554246366024, + 0.23740239441394806, + 0.10538950562477112 + ], + "credit_map": [ + 0.04777976870536804, + 0.048450518399477005, + 0.023542482405900955, + 0.013383650220930576, + 0.024478863924741745, + 0.03994946926832199, + 0.021709667518734932, + 0.025361616164445877, + 0.010512744076550007, + 0.004944439046084881, + 0.0049451482482254505, + 0.00241278437897563, + 0.004985584411770105, + 0.007349804975092411, + 0.003683150280267, + 0.005844161845743656, + 0.0130924629047513, + 0.005716010462492704, + 0.004930006805807352, + 0.0031523811630904675, + 0.006174459122121334, + 0.00814798567444086, + 0.005223993211984634, + 0.004118206910789013, + 0.02978430688381195, + 0.007667211350053549, + 0.009435881860554218, + 0.00453589903190732, + 0.011114484630525112, + 0.03163845092058182, + 0.005375730339437723, + 0.010669794864952564, + 0.02040833793580532, + 0.012047439813613892, + 0.009155763313174248, + 0.005168101750314236, + 0.008139770478010178, + 0.012975049205124378, + 0.007739044725894928, + 0.009565460495650768, + 0.005419211462140083, + 0.004350338596850634, + 0.005473752971738577, + 0.00443823728710413, + 0.0037747365422546864, + 0.0041278996504843235, + 0.0069899288937449455, + 0.0038131731562316418, + 0.003700412344187498, + 0.0029213791713118553, + 0.008348891511559486, + 0.00429347762838006, + 0.006438034120947123, + 0.006383811589330435, + 0.004961542785167694, + 0.003507734974846244, + 0.004861608613282442, + 0.0046821413561701775, + 0.004941385705024004, + 0.004560078028589487, + 0.005236209835857153, + 0.005226269364356995, + 0.010386624373495579, + 0.0030611094553023577, + 0.024625228717923164, + 0.023427898064255714, + 0.04424256458878517, + 0.04461431875824928, + 0.017033588141202927, + 0.024173995479941368, + 0.037045035511255264, + 0.022239776328206062, + 0.01634407602250576, + 0.009410851635038853, + 0.01255161315202713, + 0.0101507268846035, + 0.009729577228426933, + 0.01493515633046627, + 0.01717415452003479, + 0.01509333960711956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0036183618940413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.070022813975811 + ], + "opposite_logprob_tokens": [ + -2.6950228214263916 + ], + "logprob_reward": -0.070022813975811, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.070022813975811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22093865275382996, + 0.08590761572122574, + 0.16559448838233948, + 0.09109915047883987, + 0.12261587381362915, + 0.057779088616371155, + 0.020559517666697502, + 0.07035798579454422, + 0.06185050308704376, + 0.10329710692167282 + ], + "credit_map": [ + 0.058570653200149536, + 0.032667893916368484, + 0.02710026502609253, + 0.01473393477499485, + 0.018337097018957138, + 0.03957724571228027, + 0.014307299628853798, + 0.015644263476133347, + 0.01703735440969467, + 0.016400113701820374, + 0.012816941365599632, + 0.0037228036671876907, + 0.009762845002114773, + 0.011287286877632141, + 0.007419641595333815, + 0.007460627239197493, + 0.036468952894210815, + 0.02182679995894432, + 0.018050337210297585, + 0.009291453287005424, + 0.01995300129055977, + 0.023478049784898758, + 0.015040350146591663, + 0.02148553729057312, + 0.0271169263869524, + 0.005531309172511101, + 0.00983483623713255, + 0.005016716662794352, + 0.00658259866759181, + 0.027059273794293404, + 0.004333713091909885, + 0.0056237708777189255, + 0.029714643955230713, + 0.01909322291612625, + 0.013453603722155094, + 0.010205348953604698, + 0.01317562349140644, + 0.014687023125588894, + 0.009871210902929306, + 0.012415195815265179, + 0.013838819228112698, + 0.005586870480328798, + 0.00886551197618246, + 0.00708822812885046, + 0.005031564738601446, + 0.007272291928529739, + 0.005691887810826302, + 0.004403919912874699, + 0.0022155349142849445, + 0.0029423306696116924, + 0.0025012558326125145, + 0.0030516826082021, + 0.0025282239075750113, + 0.002589850453659892, + 0.002693507820367813, + 0.002037131693214178, + 0.009837055578827858, + 0.005080822855234146, + 0.00817838590592146, + 0.00856095366179943, + 0.008146872743964195, + 0.008260584436357021, + 0.01125939842313528, + 0.011033916845917702, + 0.007690859027206898, + 0.009182826615869999, + 0.007759324740618467, + 0.007778194732964039, + 0.007931665517389774, + 0.008017351850867271, + 0.007205065805464983, + 0.006285218521952629, + 0.02019699104130268, + 0.00818199384957552, + 0.007262093015015125, + 0.010819362476468086, + 0.00996799860149622, + 0.020010657608509064, + 0.012680104933679104, + 0.014177903532981873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.070022813975811 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, focusing on an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in the turbulent, murky, yellow-brown water. Soft, diffuse morning light penetrates the mist, illuminating his sparse, short, graying hair and equally graying beard. Drops of water trickle down his furrowed brow and cheeks strained with effort. Bare to the waist, his muscles, hardened by years of toil, now appear limp. He leans back, struggling to keep his feet steady in the muddy water beneath the surface, bracing himself against the powerful current. He clutched a rough, frayed hemp rope tightly in his hands. With every tremor of his muscles and every heavy breath, he dragged a mysterious old wooden box, partially covered in weeds and silt, inch by inch from the surging river. The box's edges, inlaid with rusted brass corners, gleamed dimly in the dim light, while icy water flowed continuously from the cracks in the box. The entire scene was filled with a sense of primal power, tense suspense, and the mysterious atmosphere of discovering the unknown. The overall color tone was cold, with a primitive and solemn texture.", + "seed": 5330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5330_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": -0.16924464665353298, + "vlm_logprob": -0.16924464665353298, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.16924464665353298, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.15925738215446472 + ], + "opposite_logprob_tokens": [ + -2.284257411956787 + ], + "logprob_reward": -0.15925738215446472, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.15925738215446472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34565404057502747, + 0.05657367780804634, + 0.20420366525650024, + 0.03117969073355198, + 0.06777600944042206, + 0.0348799005150795, + 0.03940409794449806, + 0.0988701656460762, + 0.05405453220009804, + 0.06740424036979675 + ], + "credit_map": [ + 0.06551342457532883, + 0.12784095108509064, + 0.026924828067421913, + 0.01497703604400158, + 0.028694162145256996, + 0.05255402997136116, + 0.01600724831223488, + 0.013142348267138004, + 0.008535376749932766, + 0.009876691736280918, + 0.005823921412229538, + 0.004529848229140043, + 0.00766353402286768, + 0.00869875680655241, + 0.005802125204354525, + 0.005643421784043312, + 0.06678932160139084, + 0.020235033705830574, + 0.03067181445658207, + 0.006742473691701889, + 0.016904231160879135, + 0.03682985156774521, + 0.01306877937167883, + 0.01296216156333685, + 0.004380722064524889, + 0.005437046755105257, + 0.0024806561414152384, + 0.0024072020314633846, + 0.004008423537015915, + 0.003862217767164111, + 0.0037224525585770607, + 0.004880970343947411, + 0.015788745135068893, + 0.011546696536242962, + 0.012756936252117157, + 0.006219431292265654, + 0.005744497757405043, + 0.006993461865931749, + 0.0035173920914530754, + 0.005208843387663364, + 0.005752740893512964, + 0.004985574167221785, + 0.003105154260993004, + 0.004655013792216778, + 0.003094037529081106, + 0.0037673998158425093, + 0.003600974101573229, + 0.005919007584452629, + 0.00685641635209322, + 0.005176171660423279, + 0.002997784875333309, + 0.004527128767222166, + 0.004865352064371109, + 0.007173281162977219, + 0.004302706103771925, + 0.0035052550956606865, + 0.037163570523262024, + 0.03620985150337219, + 0.003953863400965929, + 0.0040049077942967415, + 0.004806909244507551, + 0.004098762292414904, + 0.005879429168999195, + 0.0027528724167495966, + 0.011758978478610516, + 0.007481578271836042, + 0.006787443999201059, + 0.005068630911409855, + 0.006311214528977871, + 0.004377448931336403, + 0.007397978566586971, + 0.004871259443461895, + 0.012522905133664608, + 0.0073360251262784, + 0.010568240657448769, + 0.0075593735091388226, + 0.00837694201618433, + 0.008214028552174568, + 0.006269431207329035, + 0.00655729416757822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.15925738215446472 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.2521198093891144 + ], + "opposite_logprob_tokens": [ + -1.502119779586792 + ], + "logprob_reward": -0.2521198093891144, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521198093891144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09728197008371353, + 0.03646079823374748, + 0.4133656919002533, + 0.0240822434425354, + 0.033551175147295, + 0.027304144576191902, + 0.03328897804021835, + 0.1295682042837143, + 0.16175079345703125, + 0.043345920741558075 + ], + "credit_map": [ + 0.016044268384575844, + 0.01175670139491558, + 0.01352066732943058, + 0.011949826031923294, + 0.010292202234268188, + 0.009087352082133293, + 0.010267074219882488, + 0.01436387188732624, + 0.006191451568156481, + 0.007927200756967068, + 0.0033110473304986954, + 0.0023033865727484226, + 0.00374688021838665, + 0.006566636264324188, + 0.0037371560465544462, + 0.0026770399417728186, + 0.16465669870376587, + 0.02923794463276863, + 0.1064252108335495, + 0.015363725833594799, + 0.024653196334838867, + 0.04735010117292404, + 0.0074271648190915585, + 0.018251683562994003, + 0.005952585022896528, + 0.004080195911228657, + 0.00322521710768342, + 0.0013018043246120214, + 0.0026817023754119873, + 0.0022918996401131153, + 0.001878734678030014, + 0.002670103218406439, + 0.007784625515341759, + 0.005739085376262665, + 0.005576134659349918, + 0.0033172317780554295, + 0.00291883060708642, + 0.0037066100630909204, + 0.0018857879331335425, + 0.0026228714268654585, + 0.003820691956207156, + 0.0036581815220415592, + 0.0033615247812122107, + 0.0030491554643958807, + 0.004471985623240471, + 0.0028481963090598583, + 0.0033378428779542446, + 0.0027565639466047287, + 0.004648659378290176, + 0.005730763543397188, + 0.0041383178904652596, + 0.005128456745296717, + 0.003913101274520159, + 0.004101728554815054, + 0.002956909593194723, + 0.002671040827408433, + 0.05308286100625992, + 0.05077911540865898, + 0.004194148350507021, + 0.00353783811442554, + 0.005827466491609812, + 0.0039918082766234875, + 0.005036904010921717, + 0.0031180724035948515, + 0.03299413621425629, + 0.025454020127654076, + 0.025351034477353096, + 0.019945312291383743, + 0.017110364511609077, + 0.014641625806689262, + 0.016622165217995644, + 0.009632130153477192, + 0.006828187964856625, + 0.0046036806888878345, + 0.006517646834254265, + 0.005236088298261166, + 0.00618271017447114, + 0.00515170581638813, + 0.0057615311816334724, + 0.0030643651261925697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.2521198093891144 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.22560729086399078 + ], + "opposite_logprob_tokens": [ + -1.6006072759628296 + ], + "logprob_reward": -0.22560729086399078, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560729086399078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07879115641117096, + 0.12080108374357224, + 0.6281387805938721, + 0.027217362076044083, + 0.03376822918653488, + 0.02906101569533348, + 0.014228247106075287, + 0.02847868576645851, + 0.018742842599749565, + 0.020772576332092285 + ], + "credit_map": [ + 0.012605555355548859, + 0.01853068172931671, + 0.005278205499053001, + 0.006458206567913294, + 0.009644032455980778, + 0.009080642834305763, + 0.007324181962758303, + 0.009869643487036228, + 0.020838089287281036, + 0.01996404305100441, + 0.010642188601195812, + 0.008535643108189106, + 0.011782110668718815, + 0.027710573747754097, + 0.01222917903214693, + 0.009099258109927177, + 0.2469675987958908, + 0.04494158923625946, + 0.15846137702465057, + 0.022741366177797318, + 0.0365992933511734, + 0.07526273280382156, + 0.014915278181433678, + 0.028249597176909447, + 0.004776351153850555, + 0.0058133103884756565, + 0.0020964231807738543, + 0.0014662062749266624, + 0.005282963626086712, + 0.0034664610866457224, + 0.001560969976708293, + 0.002754676854237914, + 0.00533668277785182, + 0.004825342446565628, + 0.005357272922992706, + 0.0031288692262023687, + 0.0047542257234454155, + 0.005042603705078363, + 0.0032212580554187298, + 0.0021019731648266315, + 0.0067219240590929985, + 0.00656863721087575, + 0.0023554321378469467, + 0.002321284031495452, + 0.0029606858734041452, + 0.003230356378480792, + 0.002571647986769676, + 0.002331050578504801, + 0.002533740596845746, + 0.0018628144171088934, + 0.0015135920839384198, + 0.0022320488933473825, + 0.0015827888855710626, + 0.00151198694948107, + 0.0018409898038953543, + 0.0011502860579639673, + 0.010110588744282722, + 0.009782415814697742, + 0.001612342894077301, + 0.0016089084092527628, + 0.0013944742968305945, + 0.001220197882503271, + 0.0014556333189830184, + 0.0012941252207383513, + 0.002447001403197646, + 0.0021477460395544767, + 0.0033516341354697943, + 0.0020294252317398787, + 0.002093392424285412, + 0.0021905654575675726, + 0.0029245619662106037, + 0.0015585148939862847, + 0.0026259045116603374, + 0.0024592378176748753, + 0.002257270272821188, + 0.0016651403857395053, + 0.0035244436003267765, + 0.0026676449924707413, + 0.0034585886169224977, + 0.00211434718221426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.22560729086399078 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.007637230679392815 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.007637230679392815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637230679392815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1065756231546402, + 0.04369998723268509, + 0.552022397518158, + 0.03538266569375992, + 0.0374918133020401, + 0.04078742116689682, + 0.04099559038877487, + 0.06556914001703262, + 0.054847247898578644, + 0.022628095000982285 + ], + "credit_map": [ + 0.013922177255153656, + 0.01741710677742958, + 0.010993043892085552, + 0.009962804615497589, + 0.014008570462465286, + 0.012733408249914646, + 0.011492558754980564, + 0.016045957803726196, + 0.006028978619724512, + 0.008508602157235146, + 0.004166970495134592, + 0.0038049323484301567, + 0.004983614198863506, + 0.006572948768734932, + 0.004029251169413328, + 0.005604690406471491, + 0.2201189249753952, + 0.040554583072662354, + 0.14565886557102203, + 0.020116135478019714, + 0.03350409120321274, + 0.05758725479245186, + 0.011219068430364132, + 0.02326350472867489, + 0.004247152246534824, + 0.00667495746165514, + 0.0027585828211158514, + 0.0017263790359720588, + 0.007321001496165991, + 0.005187903065234423, + 0.0031666343566030264, + 0.0043000588193535805, + 0.007418110966682434, + 0.006677961442619562, + 0.005436097737401724, + 0.003359988797456026, + 0.004113906063139439, + 0.0050308131612837315, + 0.0026719376910477877, + 0.0027830004692077637, + 0.009984413161873817, + 0.008846904151141644, + 0.004522158298641443, + 0.004467302933335304, + 0.003135736333206296, + 0.0036230464465916157, + 0.002975313924252987, + 0.003232547314837575, + 0.0060905651189386845, + 0.006108677480369806, + 0.004593313671648502, + 0.004774570465087891, + 0.005061199422925711, + 0.00774815259501338, + 0.0036675541196018457, + 0.002951565198600292, + 0.020701996982097626, + 0.017301440238952637, + 0.005082256160676479, + 0.005463150329887867, + 0.004420575685799122, + 0.0033682365901768208, + 0.005330804735422134, + 0.0039006853476166725, + 0.007950021885335445, + 0.005308945197612047, + 0.005017770919948816, + 0.0033946153707802296, + 0.008209630846977234, + 0.008066988550126553, + 0.013008339330554008, + 0.0038909430149942636, + 0.0035700041335076094, + 0.002473230939358473, + 0.0022431747056543827, + 0.0023065058048814535, + 0.0035464977845549583, + 0.002928899135440588, + 0.0034490549005568027, + 0.0021107306238263845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.007637230679392815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.2016015201807022 + ], + "opposite_logprob_tokens": [ + -1.701601505279541 + ], + "logprob_reward": -0.2016015201807022, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016015201807022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3477024435997009, + 0.054245077073574066, + 0.41309237480163574, + 0.0358128622174263, + 0.02631267160177231, + 0.012657705694437027, + 0.02016594633460045, + 0.036279190331697464, + 0.03275512158870697, + 0.02097652293741703 + ], + "credit_map": [ + 0.052060116082429886, + 0.1246686726808548, + 0.034480076283216476, + 0.022651763632893562, + 0.02356361784040928, + 0.05464747175574303, + 0.015622508712112904, + 0.020008254796266556, + 0.008805694989860058, + 0.009402954950928688, + 0.005189565476030111, + 0.0033279459457844496, + 0.006919767241925001, + 0.010661454871296883, + 0.0058675711043179035, + 0.004070127848535776, + 0.16474243998527527, + 0.029514914378523827, + 0.1062377393245697, + 0.014698450453579426, + 0.025026965886354446, + 0.047384560108184814, + 0.007752840407192707, + 0.017734531313180923, + 0.005226879380643368, + 0.008908377029001713, + 0.003375413827598095, + 0.002503007650375366, + 0.005835572257637978, + 0.003610217012465, + 0.002876654267311096, + 0.0034767414908856153, + 0.006226645316928625, + 0.004455707501620054, + 0.004603886976838112, + 0.0021125932689756155, + 0.002251134254038334, + 0.0024308969732373953, + 0.00184829521458596, + 0.00238351384177804, + 0.002381925005465746, + 0.0025162920355796814, + 0.0011290688998997211, + 0.0015908816130831838, + 0.0012411063071340322, + 0.0011863360414281487, + 0.001117292675189674, + 0.001494804979301989, + 0.00306006264872849, + 0.002279856475070119, + 0.0021227002143859863, + 0.002694078255444765, + 0.0024836210068315268, + 0.0034333630464971066, + 0.0023447663988918066, + 0.0017475005006417632, + 0.014056405983865261, + 0.013362214900553226, + 0.0014209089567884803, + 0.0013068404514342546, + 0.0018976726569235325, + 0.0011325732339173555, + 0.00192946195602417, + 0.0011731163831427693, + 0.00479040015488863, + 0.0036336067132651806, + 0.0059442538768053055, + 0.00296903308480978, + 0.004299080464988947, + 0.00316763948649168, + 0.00541675416752696, + 0.0025343578308820724, + 0.0033185386564582586, + 0.002254274906590581, + 0.002698488300666213, + 0.0036124633625149727, + 0.0024215083103626966, + 0.0017713562119752169, + 0.0023813408333808184, + 0.002518555847927928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.2016015201807022 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, focusing on an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in the turbulent, murky, yellow-brown water. Soft, diffuse morning light penetrates the mist, illuminating his sparse, short, graying hair and equally graying beard. Drops of water trickle down his furrowed brow and cheeks strained with effort. Bare to the waist, his muscles, hardened by years of toil, now appear limp. He leans back, struggling to keep his feet steady in the muddy water beneath the surface, bracing himself against the powerful current. He clutched a rough, frayed hemp rope tightly in his hands. With every tremor of his muscles and every heavy breath, he dragged a mysterious old wooden box, partially covered in weeds and silt, inch by inch from the surging river. The box's edges, inlaid with rusted brass corners, gleamed dimly in the dim light, while icy water flowed continuously from the cracks in the box. The entire scene was filled with a sense of primal power, tense suspense, and the mysterious atmosphere of discovering the unknown. The overall color tone was cold, with a primitive and solemn texture.", + "seed": 5331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5331_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": -0.11226680940017104, + "vlm_logprob": -0.11226680940017104, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.11226680940017104, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.12079602479934692 + ], + "opposite_logprob_tokens": [ + -2.620795965194702 + ], + "logprob_reward": -0.12079602479934692, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.12079602479934692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10557113587856293, + 0.06834164261817932, + 0.055466875433921814, + 0.04001791030168533, + 0.08338075131177902, + 0.10137589275836945, + 0.12424423545598984, + 0.0922091156244278, + 0.19186362624168396, + 0.13752880692481995 + ], + "credit_map": [ + 0.015645867213606834, + 0.017504224553704262, + 0.009980672970414162, + 0.015035058371722698, + 0.01272566244006157, + 0.012423478998243809, + 0.01414823904633522, + 0.0081079276278615, + 0.008908744901418686, + 0.008601976558566093, + 0.007705399766564369, + 0.0066476562060415745, + 0.009203702211380005, + 0.011679304763674736, + 0.007095879875123501, + 0.008498979732394218, + 0.008870603516697884, + 0.00864511076360941, + 0.004917253740131855, + 0.008332932367920876, + 0.0095905140042305, + 0.007653596345335245, + 0.0035743070766329765, + 0.003882556688040495, + 0.007096974644809961, + 0.006717664189636707, + 0.003118470311164856, + 0.004437153227627277, + 0.005126231350004673, + 0.005248107016086578, + 0.003938696812838316, + 0.004334613215178251, + 0.015120320953428745, + 0.009587235748767853, + 0.0075810872949659824, + 0.010808272287249565, + 0.014260823838412762, + 0.007874153554439545, + 0.009370183572173119, + 0.008778671734035015, + 0.01430931594222784, + 0.011976155452430248, + 0.012332391925156116, + 0.00868313666433096, + 0.01372556947171688, + 0.015395850874483585, + 0.015717552974820137, + 0.009235921315848827, + 0.019031193107366562, + 0.011177390813827515, + 0.012098746374249458, + 0.01977360248565674, + 0.013112189248204231, + 0.017873184755444527, + 0.01884302869439125, + 0.012334895320236683, + 0.012282087467610836, + 0.009461932815611362, + 0.008533533662557602, + 0.010012121871113777, + 0.009998412802815437, + 0.010091579519212246, + 0.02463836595416069, + 0.007191080134361982, + 0.0183534137904644, + 0.019362539052963257, + 0.024639355018734932, + 0.014082219451665878, + 0.011592379771173, + 0.032870758324861526, + 0.05101601034402847, + 0.019946949556469917, + 0.014293733984231949, + 0.014644746668636799, + 0.014203001745045185, + 0.014186209067702293, + 0.013869617134332657, + 0.014282286167144775, + 0.0349966399371624, + 0.017052575945854187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.12079602479934692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.2255803644657135 + ], + "opposite_logprob_tokens": [ + -1.6005803346633911 + ], + "logprob_reward": -0.2255803644657135, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255803644657135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07745010405778885, + 0.03026515059173107, + 0.05000925809144974, + 0.029561281204223633, + 0.09326542913913727, + 0.12050145864486694, + 0.1035868301987648, + 0.13069596886634827, + 0.15175214409828186, + 0.21291233599185944 + ], + "credit_map": [ + 0.008390573784708977, + 0.010646078735589981, + 0.0079426234588027, + 0.012392286211252213, + 0.008300182409584522, + 0.008676601573824883, + 0.009852688759565353, + 0.011249067261815071, + 0.0036951005458831787, + 0.003709520446136594, + 0.0034354322124272585, + 0.002320729661732912, + 0.0034714925568550825, + 0.006417535245418549, + 0.0038241383153945208, + 0.0033912030048668385, + 0.01313120499253273, + 0.007269313558936119, + 0.0037937432061880827, + 0.0039189946837723255, + 0.00629819929599762, + 0.008849105797708035, + 0.0038562973495572805, + 0.0028923992067575455, + 0.006758027710020542, + 0.003597065107896924, + 0.002046622335910797, + 0.003584575140848756, + 0.0038856426253914833, + 0.004301257897168398, + 0.002731423592194915, + 0.0026566646993160248, + 0.014392306096851826, + 0.012764270417392254, + 0.006858911830931902, + 0.01479374524205923, + 0.013330936431884766, + 0.008159084245562553, + 0.011102508753538132, + 0.011863664723932743, + 0.014212815091013908, + 0.014827795326709747, + 0.012898299843072891, + 0.012720552273094654, + 0.017514564096927643, + 0.019393522292375565, + 0.017304779961705208, + 0.011629126034677029, + 0.013987339101731777, + 0.009664634242653847, + 0.010676276870071888, + 0.0175813976675272, + 0.011548235081136227, + 0.013182481750845909, + 0.0160553976893425, + 0.010891066864132881, + 0.020653847604990005, + 0.01932298205792904, + 0.012070546858012676, + 0.016468672081828117, + 0.014319667592644691, + 0.014533363282680511, + 0.022300662472844124, + 0.011026219464838505, + 0.019039131700992584, + 0.01607893593609333, + 0.0311010479927063, + 0.011789810843765736, + 0.009961182251572609, + 0.020933276042342186, + 0.028423847630620003, + 0.014424912631511688, + 0.04047178849577904, + 0.014239388518035412, + 0.013100341893732548, + 0.0829976424574852, + 0.02017383463680744, + 0.01562526449561119, + 0.017454100772738457, + 0.008849984034895897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.2255803644657135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.029809892177581787 + ], + "opposite_logprob_tokens": [ + -3.5298099517822266 + ], + "logprob_reward": -0.029809892177581787, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809892177581787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22640393674373627, + 0.04462407901883125, + 0.06515563279390335, + 0.05528770759701729, + 0.04373972862958908, + 0.04994183033704758, + 0.11812357604503632, + 0.191085085272789, + 0.0414457730948925, + 0.1641925871372223 + ], + "credit_map": [ + 0.02640356309711933, + 0.028248822316527367, + 0.018404483795166016, + 0.03870395943522453, + 0.02283480577170849, + 0.020442193374037743, + 0.04228933900594711, + 0.02907678298652172, + 0.006728820037096739, + 0.0043681045062839985, + 0.005723864305764437, + 0.004226914141327143, + 0.0042328257113695145, + 0.009521153755486012, + 0.004750293213874102, + 0.0050721075385808945, + 0.01632543094456196, + 0.00647629052400589, + 0.010855715721845627, + 0.0068233422935009, + 0.006839529145509005, + 0.007949553430080414, + 0.005091784056276083, + 0.004793988075107336, + 0.01702180877327919, + 0.007861071266233921, + 0.004960415884852409, + 0.004340541083365679, + 0.006085969042032957, + 0.00556983845308423, + 0.004359112121164799, + 0.00508895143866539, + 0.007027109153568745, + 0.0038217713590711355, + 0.004304731730371714, + 0.004099968355149031, + 0.007994874380528927, + 0.00456751324236393, + 0.0057508861646056175, + 0.006172872614115477, + 0.009417081251740456, + 0.006289651617407799, + 0.004581397864967585, + 0.003273612819612026, + 0.0077116177417337894, + 0.00750917149707675, + 0.007541555445641279, + 0.003617743030190468, + 0.016936613246798515, + 0.011329460889101028, + 0.01205400750041008, + 0.017070701345801353, + 0.01500889752060175, + 0.016094854101538658, + 0.018552280962467194, + 0.011076766066253185, + 0.028994997963309288, + 0.02441992051899433, + 0.022727034986019135, + 0.018993787467479706, + 0.016846204176545143, + 0.02297256700694561, + 0.03474853187799454, + 0.021382052451372147, + 0.004092071671038866, + 0.004913009237498045, + 0.003979817032814026, + 0.0034299136605113745, + 0.0030549822840839624, + 0.00617938581854105, + 0.012732032686471939, + 0.003064566059038043, + 0.028541846200823784, + 0.01044065784662962, + 0.015118660405278206, + 0.0674968957901001, + 0.017268074676394463, + 0.010857295244932175, + 0.010327247902750969, + 0.004141922574490309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.029809892177581787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.0052565219812095165 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052565219812095165, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052565219812095165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1391272395849228, + 0.03640230745077133, + 0.12355160713195801, + 0.046428512781858444, + 0.07682084292173386, + 0.052344512194395065, + 0.135861337184906, + 0.17281413078308105, + 0.08021096885204315, + 0.1364385485649109 + ], + "credit_map": [ + 0.017403777688741684, + 0.019082920625805855, + 0.012483390048146248, + 0.02100999280810356, + 0.014615249820053577, + 0.013278456404805183, + 0.023451853543519974, + 0.017801586538553238, + 0.004810862243175507, + 0.00419554952532053, + 0.003390002530068159, + 0.004027342889457941, + 0.006163288373500109, + 0.004920093342661858, + 0.004821212496608496, + 0.004073952324688435, + 0.016598282381892204, + 0.019513718783855438, + 0.008190387859940529, + 0.01633564382791519, + 0.025460895150899887, + 0.016719568520784378, + 0.01155778206884861, + 0.00917532667517662, + 0.008298185653984547, + 0.0057970392517745495, + 0.004360816907137632, + 0.005354320164769888, + 0.007669923827052116, + 0.006355845835059881, + 0.0040972582064569, + 0.004495119210332632, + 0.015046564862132072, + 0.009962321259081364, + 0.005249170120805502, + 0.009231192991137505, + 0.013564406894147396, + 0.008080835454165936, + 0.008892466314136982, + 0.00679387291893363, + 0.007136003579944372, + 0.00478786276653409, + 0.0051987250335514545, + 0.0039237262681126595, + 0.00898482371121645, + 0.006295576225966215, + 0.01025848276913166, + 0.005759309511631727, + 0.018527381122112274, + 0.011491146869957447, + 0.015225863084197044, + 0.019664080813527107, + 0.01363693829625845, + 0.02032921276986599, + 0.02186666987836361, + 0.015120040625333786, + 0.014434004202485085, + 0.014769384637475014, + 0.00900814589112997, + 0.01752565987408161, + 0.01366344653069973, + 0.023576371371746063, + 0.06752917170524597, + 0.012307938188314438, + 0.011626936495304108, + 0.009185468778014183, + 0.016957653686404228, + 0.006971989292651415, + 0.005850445479154587, + 0.009325203485786915, + 0.014931472949683666, + 0.005361797753721476, + 0.012615817598998547, + 0.015065701678395271, + 0.015563593246042728, + 0.02830350212752819, + 0.016467144712805748, + 0.013222240842878819, + 0.025380196049809456, + 0.009820341132581234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0052565219812095165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.17989124357700348 + ], + "opposite_logprob_tokens": [ + -1.8048912286758423 + ], + "logprob_reward": -0.17989124357700348, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989124357700348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12137032300233841, + 0.12285080552101135, + 0.08578628301620483, + 0.05048913508653641, + 0.08681554347276688, + 0.07035346329212189, + 0.10682930797338486, + 0.08544043451547623, + 0.06739048659801483, + 0.20267421007156372 + ], + "credit_map": [ + 0.014431864023208618, + 0.025705302134156227, + 0.007744905073195696, + 0.018497290089726448, + 0.013733653351664543, + 0.014472934417426586, + 0.015646448358893394, + 0.01113793347030878, + 0.020252345129847527, + 0.012278716079890728, + 0.010422792285680771, + 0.00917759072035551, + 0.02011154592037201, + 0.021575577557086945, + 0.01408355962485075, + 0.014948675408959389, + 0.014568690210580826, + 0.011280023492872715, + 0.00689869187772274, + 0.013164946809411049, + 0.01219150424003601, + 0.01226391363888979, + 0.006700737867504358, + 0.008717776276171207, + 0.010473774746060371, + 0.005257539916783571, + 0.005077892914414406, + 0.004561299923807383, + 0.006205682642757893, + 0.0052780210971832275, + 0.005232315976172686, + 0.008402605541050434, + 0.01632755436003208, + 0.00948448944836855, + 0.008793792687356472, + 0.010159164667129517, + 0.012856008484959602, + 0.007408491801470518, + 0.012293144129216671, + 0.009492899291217327, + 0.008540567010641098, + 0.007328948471695185, + 0.005016464274376631, + 0.004901546519249678, + 0.011330410838127136, + 0.010899445042014122, + 0.014494928531348705, + 0.007841150276362896, + 0.023442238569259644, + 0.008987583220005035, + 0.014680739492177963, + 0.009241943247616291, + 0.014205158688127995, + 0.015580362640321255, + 0.009507319889962673, + 0.0111839659512043, + 0.012295595370233059, + 0.009635924361646175, + 0.006822733674198389, + 0.010183216072618961, + 0.009106042794883251, + 0.010760228149592876, + 0.01754513382911682, + 0.009091563522815704, + 0.009418459609150887, + 0.007191620301455259, + 0.008975512348115444, + 0.004617851693183184, + 0.00600796565413475, + 0.016016652807593346, + 0.010512169450521469, + 0.004650257993489504, + 0.03480306267738342, + 0.02356564998626709, + 0.014001918025314808, + 0.0694311335682869, + 0.02294212579727173, + 0.018120206892490387, + 0.012890812940895557, + 0.006919305771589279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.17989124357700348 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly and steadily moves forward, focusing on an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in the turbulent, murky, yellow-brown water. Soft, diffuse morning light penetrates the mist, illuminating his sparse, short, graying hair and equally graying beard. Drops of water trickle down his furrowed brow and cheeks strained with effort. Bare to the waist, his muscles, hardened by years of toil, now appear limp. He leans back, struggling to keep his feet steady in the muddy water beneath the surface, bracing himself against the powerful current. He clutched a rough, frayed hemp rope tightly in his hands. With every tremor of his muscles and every heavy breath, he dragged a mysterious old wooden box, partially covered in weeds and silt, inch by inch from the surging river. The box's edges, inlaid with rusted brass corners, gleamed dimly in the dim light, while icy water flowed continuously from the cracks in the box. The entire scene was filled with a sense of primal power, tense suspense, and the mysterious atmosphere of discovering the unknown. The overall color tone was cold, with a primitive and solemn texture.", + "seed": 5332, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_7/index_1246_seed_5332_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": -0.1691338587552309, + "vlm_logprob": -0.1691338587552309, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.1691338587552309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.17159433662891388 + ], + "opposite_logprob_tokens": [ + -2.2965943813323975 + ], + "logprob_reward": -0.17159433662891388, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.17159433662891388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1836247444152832, + 0.07052434980869293, + 0.0391678661108017, + 0.06607925891876221, + 0.048864737153053284, + 0.08071660250425339, + 0.24045275151729584, + 0.09576351195573807, + 0.07618524134159088, + 0.09862098097801208 + ], + "credit_map": [ + 0.026381615549325943, + 0.023215360939502716, + 0.01932544633746147, + 0.01745494455099106, + 0.04487326368689537, + 0.023533422499895096, + 0.018933698534965515, + 0.009906995110213757, + 0.010950145311653614, + 0.007994399406015873, + 0.005605398211628199, + 0.00456699263304472, + 0.014668259769678116, + 0.011821811087429523, + 0.007198510225862265, + 0.007718832232058048, + 0.006087450310587883, + 0.004647566471248865, + 0.0035293332766741514, + 0.003358113579452038, + 0.00831426028162241, + 0.004382368177175522, + 0.004976024851202965, + 0.003872751258313656, + 0.010462864302098751, + 0.008805841207504272, + 0.007871104404330254, + 0.0055934470146894455, + 0.010461488738656044, + 0.010958069004118443, + 0.0067011695355176926, + 0.005225272383540869, + 0.005514448508620262, + 0.007092769257724285, + 0.004452907480299473, + 0.003530209418386221, + 0.010805726982653141, + 0.007171685807406902, + 0.005582677200436592, + 0.004714309237897396, + 0.023656873032450676, + 0.014028113335371017, + 0.015575064346194267, + 0.003439453197643161, + 0.006456529255956411, + 0.00821954570710659, + 0.005549881141632795, + 0.0037911429535597563, + 0.07104998826980591, + 0.015644259750843048, + 0.01107793115079403, + 0.004841791931539774, + 0.01561473123729229, + 0.09454013407230377, + 0.017749575898051262, + 0.00993433315306902, + 0.010083113797008991, + 0.007939773611724377, + 0.02461852878332138, + 0.004611439071595669, + 0.006409919820725918, + 0.029888790100812912, + 0.007266646716743708, + 0.004945292137563229, + 0.011209474876523018, + 0.0063881948590278625, + 0.011864729225635529, + 0.012574099004268646, + 0.008955534547567368, + 0.01099362038075924, + 0.006457376293838024, + 0.007742216344922781, + 0.03513266146183014, + 0.0077594611793756485, + 0.025783486664295197, + 0.007702270057052374, + 0.0061034695245325565, + 0.00873254518955946, + 0.004438674543052912, + 0.002968410961329937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.17159433662891388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.38707470893859863 + ], + "opposite_logprob_tokens": [ + -1.1370747089385986 + ], + "logprob_reward": -0.38707470893859863, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38707470893859863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1432032734155655, + 0.06763245910406113, + 0.0458519384264946, + 0.09064500033855438, + 0.15001356601715088, + 0.08606908470392227, + 0.03560274466872215, + 0.2499406933784485, + 0.06534741818904877, + 0.06569388508796692 + ], + "credit_map": [ + 0.022221563383936882, + 0.025464454665780067, + 0.013800847344100475, + 0.01261911354959011, + 0.03171152248978615, + 0.015310312621295452, + 0.013973641209304333, + 0.008101806044578552, + 0.009801403619349003, + 0.007691888138651848, + 0.005345741752535105, + 0.005006894469261169, + 0.01400700118392706, + 0.01156502403318882, + 0.007651013322174549, + 0.006563487928360701, + 0.008326304145157337, + 0.006125571206212044, + 0.004834878258407116, + 0.004462996032088995, + 0.007513256277889013, + 0.005261207930743694, + 0.004944321699440479, + 0.004383396357297897, + 0.013027321547269821, + 0.0113303754478693, + 0.00584920309484005, + 0.008124555461108685, + 0.01675502210855484, + 0.018520405516028404, + 0.008136270567774773, + 0.008901836350560188, + 0.015224983915686607, + 0.018841559067368507, + 0.013631479814648628, + 0.00829402543604374, + 0.038755424320697784, + 0.017771374434232712, + 0.023219935595989227, + 0.014274770393967628, + 0.02008933015167713, + 0.018808819353580475, + 0.004299967084079981, + 0.004896747879683971, + 0.009249497205018997, + 0.014897491782903671, + 0.00851029809564352, + 0.005316916387528181, + 0.005777002777904272, + 0.0045451694168150425, + 0.0030366177670657635, + 0.0021509130019694567, + 0.004415358882397413, + 0.007083974778652191, + 0.0059895687736570835, + 0.0026041341479867697, + 0.02834988199174404, + 0.014270239509642124, + 0.07058725506067276, + 0.009793722070753574, + 0.009206719696521759, + 0.08858976513147354, + 0.016216933727264404, + 0.012926146388053894, + 0.011601115576922894, + 0.006495602894574404, + 0.009204806759953499, + 0.006128173321485519, + 0.004913350101560354, + 0.01751752384006977, + 0.004467010032385588, + 0.005019831471145153, + 0.019484296441078186, + 0.007279124576598406, + 0.01096658781170845, + 0.0037590956781059504, + 0.010396024212241173, + 0.005541533697396517, + 0.004524562507867813, + 0.0037426529452204704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.38707470893859863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.10034245252609253 + ], + "opposite_logprob_tokens": [ + -2.3503425121307373 + ], + "logprob_reward": -0.10034245252609253, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034245252609253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15229317545890808, + 0.04326721280813217, + 0.09664833545684814, + 0.14002707600593567, + 0.03769206255674362, + 0.14938852190971375, + 0.03895125910639763, + 0.09744947403669357, + 0.08624285459518433, + 0.15803995728492737 + ], + "credit_map": [ + 0.020712722092866898, + 0.019955461844801903, + 0.015080844983458519, + 0.011285015381872654, + 0.0363791324198246, + 0.01943628117442131, + 0.018940391018986702, + 0.01050332561135292, + 0.007134080398827791, + 0.00573904300108552, + 0.004585344344377518, + 0.0032888944260776043, + 0.007531116250902414, + 0.006150390021502972, + 0.0038140735123306513, + 0.005024272482842207, + 0.020041339099407196, + 0.013186554424464703, + 0.01242511160671711, + 0.01148032397031784, + 0.012129419483244419, + 0.012145312502980232, + 0.008988753892481327, + 0.00625152513384819, + 0.019107770174741745, + 0.018400806933641434, + 0.021220725029706955, + 0.012957751750946045, + 0.023891977965831757, + 0.01686318777501583, + 0.014714136719703674, + 0.01287071779370308, + 0.004492076579481363, + 0.005328675732016563, + 0.004541677422821522, + 0.002136204857379198, + 0.0072388495318591595, + 0.007445786613970995, + 0.003744622925296426, + 0.0027641647029668093, + 0.03430500999093056, + 0.02927316725254059, + 0.0150735042989254, + 0.008977682329714298, + 0.016876589506864548, + 0.0205294881016016, + 0.014904467388987541, + 0.00944861862808466, + 0.008030968718230724, + 0.0037344698794186115, + 0.003438372863456607, + 0.001902577350847423, + 0.004404326900839806, + 0.005902822129428387, + 0.007960923947393894, + 0.0035767999943345785, + 0.011265994980931282, + 0.00519835390150547, + 0.025668535381555557, + 0.004051098134368658, + 0.004493035841733217, + 0.03384755551815033, + 0.0073100202716887, + 0.005614886060357094, + 0.013447633944451809, + 0.00695339310914278, + 0.014192907139658928, + 0.01028952281922102, + 0.007950074970722198, + 0.019902700558304787, + 0.006521862931549549, + 0.0069847614504396915, + 0.05717042461037636, + 0.011754747480154037, + 0.04151991754770279, + 0.011726039461791515, + 0.007622126489877701, + 0.014494993723928928, + 0.00793360359966755, + 0.005818091332912445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.10034245252609253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11547207832336426, + 0.031736817210912704, + 0.03788121044635773, + 0.026976993307471275, + 0.026663372293114662, + 0.06003192439675331, + 0.11315488815307617, + 0.5085422396659851, + 0.03898627683520317, + 0.04055425524711609 + ], + "credit_map": [ + 0.015079610981047153, + 0.01628425344824791, + 0.01237487606704235, + 0.013537687249481678, + 0.023377953097224236, + 0.013701082207262516, + 0.01245051808655262, + 0.00866609439253807, + 0.005384653806686401, + 0.004054590128362179, + 0.003305975813418627, + 0.0030058634001761675, + 0.005776999983936548, + 0.004505214747041464, + 0.0031789704225957394, + 0.002524546580389142, + 0.004694838542491198, + 0.006057530175894499, + 0.003028462640941143, + 0.00216841883957386, + 0.008721460588276386, + 0.005560767371207476, + 0.003871733322739601, + 0.003777997102588415, + 0.0030074845999479294, + 0.0035344732459634542, + 0.00237902719527483, + 0.0034394378308206797, + 0.004708821885287762, + 0.003976406995207071, + 0.0032800803892314434, + 0.0026512595359236, + 0.003932393621653318, + 0.0041943201795220375, + 0.00256132404319942, + 0.002614252967759967, + 0.004995840135961771, + 0.0035998811945319176, + 0.0028734891675412655, + 0.0018918690038844943, + 0.016891926527023315, + 0.00840086117386818, + 0.009667351841926575, + 0.002697305055335164, + 0.00833372212946415, + 0.006119567435234785, + 0.005084104370325804, + 0.002837080741301179, + 0.029274826869368553, + 0.01198198925703764, + 0.004585220478475094, + 0.004010703880339861, + 0.0114455446600914, + 0.03246467933058739, + 0.013736945576965809, + 0.005654972046613693, + 0.07080177217721939, + 0.02886638604104519, + 0.15730905532836914, + 0.01851055957376957, + 0.019514190033078194, + 0.15842972695827484, + 0.02909635193645954, + 0.02601414918899536, + 0.006213534623384476, + 0.0035713831894099712, + 0.003911477513611317, + 0.0035804950166493654, + 0.0037562819197773933, + 0.00964452512562275, + 0.003192993812263012, + 0.005115585867315531, + 0.011172318831086159, + 0.004813771229237318, + 0.006817368790507317, + 0.0032120696268975735, + 0.007028793916106224, + 0.002908519236370921, + 0.00249345600605011, + 0.002107953652739525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 filmed from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving slowly and steadily forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old fisherman with deeply wrinkled, bronzed skin standing waist-deep in turbulent, murky, yellow-brown water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fisherman bare to the waist, leaning back while clutching a rough, frayed hemp rope and dragging a mysterious old wooden box with rusted brass corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall color tone cold with a primitive and solemn 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": [ + -0.17991414666175842 + ], + "opposite_logprob_tokens": [ + -1.804914116859436 + ], + "logprob_reward": -0.17991414666175842, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991414666175842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07300607860088348, + 0.02570861577987671, + 0.023506660014390945, + 0.025923533365130424, + 0.019982289522886276, + 0.06235400214791298, + 0.07380590587854385, + 0.5281357765197754, + 0.01998143270611763, + 0.14759568870067596 + ], + "credit_map": [ + 0.015306197106838226, + 0.0117570785805583, + 0.006689161527901888, + 0.006333245430141687, + 0.012531004846096039, + 0.010035061277449131, + 0.006480887532234192, + 0.0038734343834221363, + 0.004131784662604332, + 0.003030316438525915, + 0.0026575839146971703, + 0.0018404595321044326, + 0.004159658681601286, + 0.004379148595035076, + 0.0031612347811460495, + 0.0023484318517148495, + 0.00457778200507164, + 0.0028277740348130465, + 0.0028322681318968534, + 0.0027909339405596256, + 0.003470344701781869, + 0.0026559459511190653, + 0.002697610529139638, + 0.0016540002543479204, + 0.0036362772807478905, + 0.0037135810125619173, + 0.0032854750752449036, + 0.0026299995370209217, + 0.003928523976355791, + 0.003762617940083146, + 0.002675340510904789, + 0.0022917191963642836, + 0.0029201109427958727, + 0.002694325987249613, + 0.0023763396311551332, + 0.0009943223558366299, + 0.0035720360465347767, + 0.0035983524285256863, + 0.002312475349754095, + 0.0015143270138651133, + 0.019615409895777702, + 0.01095094345510006, + 0.009208346717059612, + 0.0024343726690858603, + 0.006446242798119783, + 0.007574308197945356, + 0.0035868810955435038, + 0.002537495456635952, + 0.022885426878929138, + 0.007177216466516256, + 0.0029307445511221886, + 0.0014266592916101217, + 0.00445335078984499, + 0.02756170555949211, + 0.004543827380985022, + 0.002826977986842394, + 0.06262243539094925, + 0.025676049292087555, + 0.15882602334022522, + 0.01786777749657631, + 0.016915136948227882, + 0.19030329585075378, + 0.02964956685900688, + 0.026275478303432465, + 0.005786586087197065, + 0.0022689318284392357, + 0.0029244604520499706, + 0.0013976613990962505, + 0.0016348151257261634, + 0.0037170129362493753, + 0.0014490671455860138, + 0.0008028973243199289, + 0.02073666825890541, + 0.015599251724779606, + 0.023114101961255074, + 0.009940428659319878, + 0.01823471300303936, + 0.03186284378170967, + 0.015591472387313843, + 0.012516220100224018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.17991414666175842 + } + ], + "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/0000110/group_8/index_tensor([1377, 1377])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_8/index_tensor([1377, 1377])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..07646489a716a89b781e92a24c205ab34191d094 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_8/index_tensor([1377, 1377])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a slow, majestic, wide-angle, wraparound shot, we gaze from a godlike perspective upon a vast protoplanetary disk, engaged in a magnificent gyroscope, dominating the entire expanse of the horizon. This vast cosmic vortex, composed of countless clouds of deep purple and fiery orange gas and wispy dust lanes, rotates steadily and leisurely within the blackness of space. Pulled inexorably by the powerful gravity of a central, immature protostar, the flow of matter spirals inexorably inward, forming vast gravitational vortices and denser clumps of matter. In this process, countless dust grains collide and coalesce, emitting bursts of nascent light, like countless flickering embers scattered across the velvety curtain of space. These faint, warm flashes illuminate the dark regions within the disk, heralding the birth and formation of a new world from chaos. The background of the entire picture is the deep interstellar space, dotted with distant and cold stars. The overall color tone transitions between the warm incandescent light in the center and the cold blue tones at the outer edge, creating a tranquil atmosphere that is both magnificent and full of hope for life, perfectly fitting a humanistic documentary's poetic exploration of the origin of the universe.", + "seed": 29734, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_8/index_1377_seed_29734_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": -0.05596404243260622, + "vlm_logprob": -0.05596404243260622, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.05596404243260622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01263247150927782 + ], + "opposite_logprob_tokens": [ + -4.637632369995117 + ], + "logprob_reward": -0.01263247150927782, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01263247150927782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1135086789727211, + 0.06782092154026031, + 0.05940783768892288, + 0.04355890303850174, + 0.06253433972597122, + 0.07248686254024506, + 0.3381340205669403, + 0.07910406589508057, + 0.1301933079957962, + 0.03325103968381882 + ], + "credit_map": [ + 0.010400934144854546, + 0.012063373811542988, + 0.015686823055148125, + 0.014360606670379639, + 0.012654704973101616, + 0.01468547061085701, + 0.013690231367945671, + 0.01996653899550438, + 0.007134936284273863, + 0.005490466021001339, + 0.007621974218636751, + 0.006463985424488783, + 0.012823077850043774, + 0.009769397787749767, + 0.010075386613607407, + 0.008441700600087643, + 0.006107653956860304, + 0.003970710560679436, + 0.007595793344080448, + 0.007880655117332935, + 0.009380715899169445, + 0.010226045735180378, + 0.007746601477265358, + 0.006499661132693291, + 0.005408265627920628, + 0.003833848750218749, + 0.005290011875331402, + 0.006452830508351326, + 0.005759909749031067, + 0.00442124018445611, + 0.005354700144380331, + 0.007038091775029898, + 0.0062575871124863625, + 0.004871752578765154, + 0.007206606678664684, + 0.007910219952464104, + 0.01115955039858818, + 0.005287526175379753, + 0.007148729171603918, + 0.01269236858934164, + 0.005982859525829554, + 0.007644212804734707, + 0.007907072082161903, + 0.010876908898353577, + 0.009909514337778091, + 0.007395878899842501, + 0.009696342051029205, + 0.013074072077870369, + 0.015126477926969528, + 0.02557429112493992, + 0.04106513038277626, + 0.12376872450113297, + 0.014059937559068203, + 0.07429711520671844, + 0.019001144915819168, + 0.025241201743483543, + 0.005386049393564463, + 0.005667040590196848, + 0.012515414506196976, + 0.011909585446119308, + 0.007521619088947773, + 0.009190280921757221, + 0.011165459640324116, + 0.01574861817061901, + 0.012544256635010242, + 0.008779589086771011, + 0.018174393102526665, + 0.012852253392338753, + 0.020542612299323082, + 0.0176627766340971, + 0.014663423411548138, + 0.02497401088476181, + 0.0022786487825214863, + 0.002608533948659897, + 0.003932494204491377, + 0.0023486099671572447, + 0.002332647331058979, + 0.0035520452074706554, + 0.005627994891256094, + 0.01057006511837244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.01263247150927782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026379402726888657 + ], + "opposite_logprob_tokens": [ + -3.651379346847534 + ], + "logprob_reward": -0.026379402726888657, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026379402726888657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13981282711029053, + 0.1515415459871292, + 0.06806088984012604, + 0.05421494320034981, + 0.06934477388858795, + 0.14291498064994812, + 0.10238851606845856, + 0.15180133283138275, + 0.04368118569254875, + 0.0762389525771141 + ], + "credit_map": [ + 0.01284253504127264, + 0.016314228996634483, + 0.023028133437037468, + 0.020224623382091522, + 0.013577803038060665, + 0.01736060529947281, + 0.014503855258226395, + 0.021961046382784843, + 0.020165808498859406, + 0.01131395809352398, + 0.01487867720425129, + 0.01402810774743557, + 0.028902418911457062, + 0.022510066628456116, + 0.020557517185807228, + 0.01918499357998371, + 0.008023249916732311, + 0.004950816743075848, + 0.008773406967520714, + 0.010771465487778187, + 0.00862663984298706, + 0.010703995823860168, + 0.00830794870853424, + 0.00790336448699236, + 0.006064082030206919, + 0.004185279365628958, + 0.006302789784967899, + 0.004851402714848518, + 0.008037586696445942, + 0.005169740412384272, + 0.010138983838260174, + 0.009465075097978115, + 0.007197800558060408, + 0.004670003429055214, + 0.00675168028101325, + 0.005571211222559214, + 0.013010533526539803, + 0.013297916390001774, + 0.007435687817633152, + 0.011409944854676723, + 0.009656262584030628, + 0.008378248661756516, + 0.010181981138885021, + 0.012128173373639584, + 0.011820443905889988, + 0.03465571627020836, + 0.008588749915361404, + 0.04750540107488632, + 0.006886872462928295, + 0.010626934468746185, + 0.014432314783334732, + 0.026844898238778114, + 0.00904571171849966, + 0.009187896735966206, + 0.009151460602879524, + 0.016212422400712967, + 0.009128902107477188, + 0.007884997874498367, + 0.029932739213109016, + 0.01589415781199932, + 0.013985564932227135, + 0.018537217751145363, + 0.01637999154627323, + 0.04005776345729828, + 0.006236641202121973, + 0.0033027129247784615, + 0.006653407122939825, + 0.005282500758767128, + 0.005673412699252367, + 0.00498784938827157, + 0.005370037630200386, + 0.006174623500555754, + 0.008582054637372494, + 0.005036107264459133, + 0.0106651084497571, + 0.005886730272322893, + 0.004730339162051678, + 0.008983109146356583, + 0.011091681197285652, + 0.02126382291316986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.026379402726888657 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495971441268921 + ], + "opposite_logprob_tokens": [ + -2.929959774017334 + ], + "logprob_reward": -0.05495971441268921, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495971441268921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2564954459667206, + 0.03405074402689934, + 0.05039948225021362, + 0.0349869430065155, + 0.06720790266990662, + 0.050955768674612045, + 0.05626478046178818, + 0.1728406399488449, + 0.06002078577876091, + 0.2167774736881256 + ], + "credit_map": [ + 0.025014596059918404, + 0.018191397190093994, + 0.01868031546473503, + 0.020794188603758812, + 0.027871638536453247, + 0.054348841309547424, + 0.02823532372713089, + 0.06335916370153427, + 0.0032383680809289217, + 0.0030480786226689816, + 0.004303183872252703, + 0.005046024452894926, + 0.005587659776210785, + 0.0045855119824409485, + 0.003760186256840825, + 0.00448173051699996, + 0.005611287895590067, + 0.003980312030762434, + 0.005989873316138983, + 0.0075401561334729195, + 0.005879742559045553, + 0.007785540074110031, + 0.007312835659831762, + 0.006299732252955437, + 0.003926500212401152, + 0.003544050734490156, + 0.005387157667428255, + 0.004837561398744583, + 0.004250416066497564, + 0.003989489749073982, + 0.004544055089354515, + 0.004507715813815594, + 0.006674580741673708, + 0.0037615231703966856, + 0.005065910518169403, + 0.005539081525057554, + 0.01239737682044506, + 0.014898581430315971, + 0.008854561485350132, + 0.01001629140228033, + 0.004590826574712992, + 0.005661981180310249, + 0.005400326568633318, + 0.006167665123939514, + 0.0068451594561338425, + 0.007997534237802029, + 0.0045285350643098354, + 0.009763740003108978, + 0.005037605296820402, + 0.0037466981448233128, + 0.00609110202640295, + 0.007639131508767605, + 0.004846710246056318, + 0.0137030566111207, + 0.005683800671249628, + 0.009516675025224686, + 0.009203976020216942, + 0.0068953996524214745, + 0.04657486453652382, + 0.012722622603178024, + 0.00965567585080862, + 0.021578887477517128, + 0.014486883766949177, + 0.051722340285778046, + 0.00606151856482029, + 0.003895852714776993, + 0.008732505142688751, + 0.006828680634498596, + 0.0072280894964933395, + 0.008429300040006638, + 0.008489626459777355, + 0.010355210863053799, + 0.023293275386095047, + 0.009951571933925152, + 0.027364427223801613, + 0.010314887389540672, + 0.01029740460216999, + 0.02200126089155674, + 0.042806629091501236, + 0.07074800878763199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.05495971441268921 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304775595664978 + ], + "opposite_logprob_tokens": [ + -3.1680476665496826 + ], + "logprob_reward": -0.04304775595664978, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304775595664978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30122110247612, + 0.1417582780122757, + 0.07578757405281067, + 0.038153890520334244, + 0.25924748182296753, + 0.027895445004105568, + 0.03809811547398567, + 0.058193668723106384, + 0.027735456824302673, + 0.03190898895263672 + ], + "credit_map": [ + 0.03017398901283741, + 0.023958798497915268, + 0.021994540467858315, + 0.025171935558319092, + 0.036046721041202545, + 0.057257380336523056, + 0.03691612929105759, + 0.06970158219337463, + 0.015045755542814732, + 0.012083001434803009, + 0.012577567249536514, + 0.013647709973156452, + 0.030785923823714256, + 0.023704007267951965, + 0.018581721931695938, + 0.015332587994635105, + 0.009884064085781574, + 0.0049577876925468445, + 0.008573081344366074, + 0.009487496688961983, + 0.008669765666127205, + 0.011997883208096027, + 0.012017585337162018, + 0.010199897922575474, + 0.0043258061632514, + 0.00292546721175313, + 0.003978224005550146, + 0.0032313449773937464, + 0.005582586862146854, + 0.006224635522812605, + 0.005384351592510939, + 0.006501470226794481, + 0.025027813389897346, + 0.011999286711215973, + 0.020515963435173035, + 0.01773880422115326, + 0.04180052876472473, + 0.06540948897600174, + 0.03373357653617859, + 0.04302200675010681, + 0.0031459361780434847, + 0.0017411871813237667, + 0.002496397588402033, + 0.0031669829040765762, + 0.0032532408367842436, + 0.004864134360104799, + 0.0022320898715406656, + 0.006995473988354206, + 0.0022949478588998318, + 0.0029298115987330675, + 0.005472078453749418, + 0.014977769926190376, + 0.0021087261848151684, + 0.004846180323511362, + 0.002231546677649021, + 0.0032370504923164845, + 0.0031793646048754454, + 0.002539413282647729, + 0.014062166213989258, + 0.004600823856890202, + 0.0033391891047358513, + 0.007732072379440069, + 0.004779687616974115, + 0.01796095073223114, + 0.0029828313272446394, + 0.0018988930387422442, + 0.003853259840980172, + 0.0031872487161308527, + 0.003352804807946086, + 0.0039564925245940685, + 0.0039012201596051455, + 0.0046027060598134995, + 0.004109279252588749, + 0.0019097008043900132, + 0.004293931648135185, + 0.0014086189912632108, + 0.0023111114278435707, + 0.002983034821227193, + 0.006037633400410414, + 0.008855675347149372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04304775595664978 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280086755752563 + ], + "opposite_logprob_tokens": [ + -2.017800807952881 + ], + "logprob_reward": -0.14280086755752563, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280086755752563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3765542805194855, + 0.049326151609420776, + 0.079489566385746, + 0.026240792125463486, + 0.0704042986035347, + 0.05748909339308739, + 0.07074849307537079, + 0.045008253306150436, + 0.06833897531032562, + 0.15640011429786682 + ], + "credit_map": [ + 0.0534806065261364, + 0.038590118288993835, + 0.0325513556599617, + 0.037128996104002, + 0.05099926516413689, + 0.05113754794001579, + 0.059118133038282394, + 0.053548261523246765, + 0.0050677149556577206, + 0.0038856216706335545, + 0.004870915785431862, + 0.0075833601877093315, + 0.00781888049095869, + 0.006789592560380697, + 0.005727140232920647, + 0.007582928519695997, + 0.009620154276490211, + 0.00533510884270072, + 0.010385764762759209, + 0.010789514519274235, + 0.011243791319429874, + 0.012096342630684376, + 0.01009289175271988, + 0.00992599967867136, + 0.002566036768257618, + 0.0023379239719361067, + 0.003722158260643482, + 0.0031513895373791456, + 0.0039741829968988895, + 0.0032359932083636522, + 0.00315475114621222, + 0.004098356701433659, + 0.006981512065976858, + 0.0037513887509703636, + 0.006679257843643427, + 0.004295199643820524, + 0.012966126203536987, + 0.015295098535716534, + 0.007277201861143112, + 0.01315851416438818, + 0.004611475393176079, + 0.003458732971921563, + 0.006564071401953697, + 0.007255980279296637, + 0.0052469652146101, + 0.009570292197167873, + 0.0057881977409124374, + 0.014993380755186081, + 0.004471328109502792, + 0.006196871865540743, + 0.007973549887537956, + 0.01937182806432247, + 0.006194287445396185, + 0.012848787009716034, + 0.006529504898935556, + 0.007162333000451326, + 0.003704165108501911, + 0.003962207585573196, + 0.009277782402932644, + 0.008295869454741478, + 0.0049158865585923195, + 0.004587133415043354, + 0.005128704477101564, + 0.005136503372341394, + 0.006265066564083099, + 0.0042574601247906685, + 0.010111801326274872, + 0.006790847051888704, + 0.009948646649718285, + 0.010405220091342926, + 0.00857192650437355, + 0.011988007463514805, + 0.017992736771702766, + 0.006461014039814472, + 0.020147519186139107, + 0.0075185904279351234, + 0.006914685945957899, + 0.01555757224559784, + 0.032032329589128494, + 0.04977566376328468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.14280086755752563 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow, majestic, wide-angle, wraparound shot, we gaze from a godlike perspective upon a vast protoplanetary disk, engaged in a magnificent gyroscope, dominating the entire expanse of the horizon. This vast cosmic vortex, composed of countless clouds of deep purple and fiery orange gas and wispy dust lanes, rotates steadily and leisurely within the blackness of space. Pulled inexorably by the powerful gravity of a central, immature protostar, the flow of matter spirals inexorably inward, forming vast gravitational vortices and denser clumps of matter. In this process, countless dust grains collide and coalesce, emitting bursts of nascent light, like countless flickering embers scattered across the velvety curtain of space. These faint, warm flashes illuminate the dark regions within the disk, heralding the birth and formation of a new world from chaos. The background of the entire picture is the deep interstellar space, dotted with distant and cold stars. The overall color tone transitions between the warm incandescent light in the center and the cold blue tones at the outer edge, creating a tranquil atmosphere that is both magnificent and full of hope for life, perfectly fitting a humanistic documentary's poetic exploration of the origin of the universe.", + "seed": 29735, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_8/index_1377_seed_29735_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": -0.15078646168112755, + "vlm_logprob": -0.15078646168112755, + "vlm_yes_no_margin": 2.324999988079071, + "vlm_token_credit": -0.15078646168112755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021039865911006927 + ], + "opposite_logprob_tokens": [ + -4.021039962768555 + ], + "logprob_reward": -0.021039865911006927, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021039865911006927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09449931234121323, + 0.15030443668365479, + 0.16673296689987183, + 0.09010426700115204, + 0.1031627207994461, + 0.1561829000711441, + 0.06849505752325058, + 0.08224796503782272, + 0.029588842764496803, + 0.05868151783943176 + ], + "credit_map": [ + 0.011402285657823086, + 0.012657254934310913, + 0.012579992413520813, + 0.010921201668679714, + 0.011948167346417904, + 0.009756057523190975, + 0.009876076132059097, + 0.015358269214630127, + 0.01435130462050438, + 0.009924896992743015, + 0.04530021920800209, + 0.017195584252476692, + 0.008476161397993565, + 0.01060200110077858, + 0.01114282850176096, + 0.03331143781542778, + 0.018567267805337906, + 0.012495835311710835, + 0.007642506621778011, + 0.008291065692901611, + 0.012179972603917122, + 0.048395946621894836, + 0.011400997638702393, + 0.04775936156511307, + 0.011392808519303799, + 0.01015055738389492, + 0.012810208834707737, + 0.009681967087090015, + 0.01088687777519226, + 0.009438826702535152, + 0.013331440277397633, + 0.012411578558385372, + 0.009311075322329998, + 0.011239803396165371, + 0.013893219642341137, + 0.012363400310277939, + 0.011178875342011452, + 0.010365980677306652, + 0.014836931601166725, + 0.019973430782556534, + 0.022294387221336365, + 0.009717267006635666, + 0.023509876802563667, + 0.011526189744472504, + 0.010869286954402924, + 0.01773247867822647, + 0.01919396035373211, + 0.0413394533097744, + 0.00796882901340723, + 0.006132515612989664, + 0.015867307782173157, + 0.007010804023593664, + 0.0065360707230865955, + 0.008344932459294796, + 0.008352345786988735, + 0.008282246068120003, + 0.011929336935281754, + 0.005671164020895958, + 0.011066215112805367, + 0.008098011836409569, + 0.009950228035449982, + 0.014401973225176334, + 0.011731421574950218, + 0.009399614296853542, + 0.002776175271719694, + 0.0023421382065862417, + 0.0056255836971104145, + 0.0031854293774813414, + 0.003668837249279022, + 0.003134462982416153, + 0.004308524541556835, + 0.004547689575701952, + 0.0068436977453529835, + 0.005187974311411381, + 0.009094619192183018, + 0.008745732717216015, + 0.005559348035603762, + 0.006247914861887693, + 0.006614810787141323, + 0.010387417860329151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.021039865911006927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04310370609164238 + ], + "opposite_logprob_tokens": [ + -3.1681036949157715 + ], + "logprob_reward": -0.04310370609164238, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04310370609164238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09686292707920074, + 0.20685994625091553, + 0.07713249325752258, + 0.1649671196937561, + 0.06547942757606506, + 0.044682975858449936, + 0.1382896900177002, + 0.12615859508514404, + 0.020296834409236908, + 0.05926999822258949 + ], + "credit_map": [ + 0.009443074464797974, + 0.009877811186015606, + 0.021593928337097168, + 0.009226907975971699, + 0.009994439780712128, + 0.007827707566320896, + 0.014855263754725456, + 0.014043794013559818, + 0.021179378032684326, + 0.011208850890398026, + 0.06751971691846848, + 0.02207065187394619, + 0.009239356964826584, + 0.013777002692222595, + 0.014842614531517029, + 0.047022365033626556, + 0.009970465674996376, + 0.006137375254184008, + 0.003508134512230754, + 0.0036090987268835306, + 0.005351081024855375, + 0.021938014775514603, + 0.0053145429119467735, + 0.021303774788975716, + 0.012694288976490498, + 0.011086557060480118, + 0.06130494922399521, + 0.02111080102622509, + 0.010184591636061668, + 0.023636050522327423, + 0.011355054564774036, + 0.013594808988273144, + 0.005269992630928755, + 0.007540042046457529, + 0.010105311870574951, + 0.007961471565067768, + 0.006721280515193939, + 0.006593743804842234, + 0.009196796454489231, + 0.012090783566236496, + 0.005373230203986168, + 0.003944440744817257, + 0.005591865628957748, + 0.004254499915987253, + 0.004680273588746786, + 0.004742184653878212, + 0.0059440853074193, + 0.010152394883334637, + 0.014632516540586948, + 0.010371307842433453, + 0.02981342375278473, + 0.012283986434340477, + 0.015412148088216782, + 0.01804306171834469, + 0.016681339591741562, + 0.02105189673602581, + 0.01157782319933176, + 0.00813211314380169, + 0.01732715778052807, + 0.018025347962975502, + 0.01636476442217827, + 0.017393434420228004, + 0.025044448673725128, + 0.012293498031795025, + 0.0015005519380792975, + 0.001555013470351696, + 0.00458788126707077, + 0.0018450835486873984, + 0.0016258024843409657, + 0.0025111553259193897, + 0.002160398056730628, + 0.004510947968810797, + 0.0070595345459878445, + 0.004392355214804411, + 0.009572092443704605, + 0.008809438906610012, + 0.0052810246124863625, + 0.006799946539103985, + 0.007986561395227909, + 0.009369036182761192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.04310370609164238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903924375772476 + ], + "opposite_logprob_tokens": [ + -2.4640393257141113 + ], + "logprob_reward": -0.08903924375772476, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903924375772476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10600468516349792, + 0.052315011620521545, + 0.09660916775465012, + 0.31936508417129517, + 0.10176072269678116, + 0.04487961530685425, + 0.11617010831832886, + 0.09459993988275528, + 0.04907463863492012, + 0.019221074879169464 + ], + "credit_map": [ + 0.009632157161831856, + 0.015676816925406456, + 0.015856588259339333, + 0.008675298653542995, + 0.01417480781674385, + 0.010390681214630604, + 0.013174686580896378, + 0.018423642963171005, + 0.008531340397894382, + 0.007187813054770231, + 0.007465487811714411, + 0.006718779448419809, + 0.004622501786798239, + 0.004318276885896921, + 0.00527418777346611, + 0.008196624927222729, + 0.013538667000830173, + 0.007788036484271288, + 0.005491116549819708, + 0.0046387105248868465, + 0.0073876939713954926, + 0.024378281086683273, + 0.007749834563583136, + 0.025636820122599602, + 0.021115530282258987, + 0.018628794699907303, + 0.11253978312015533, + 0.0402984693646431, + 0.02034936472773552, + 0.06298099458217621, + 0.021742623299360275, + 0.021709522232413292, + 0.010183273814618587, + 0.015126136131584644, + 0.012519219890236855, + 0.016917582601308823, + 0.013546458445489407, + 0.008591593243181705, + 0.013398121111094952, + 0.011478334665298462, + 0.0037085481453686953, + 0.004934687167406082, + 0.006764227990061045, + 0.0066490136086940765, + 0.005729015450924635, + 0.003856707364320755, + 0.007126621901988983, + 0.006110788322985172, + 0.013192077167332172, + 0.011212051846086979, + 0.02523670345544815, + 0.011762460693717003, + 0.011554323136806488, + 0.012151489034295082, + 0.014441032893955708, + 0.016619965434074402, + 0.013974165543913841, + 0.006500197574496269, + 0.014520031400024891, + 0.007465173490345478, + 0.010702974162995815, + 0.017974480986595154, + 0.01052796933799982, + 0.012934946455061436, + 0.006198294926434755, + 0.004705472383648157, + 0.006642954424023628, + 0.0059152087196707726, + 0.005430576857179403, + 0.007383112795650959, + 0.006205754820257425, + 0.0065932609140872955, + 0.001983643975108862, + 0.0020887318532913923, + 0.003234913107007742, + 0.0024902529548853636, + 0.0014540075790137053, + 0.0019340348662808537, + 0.0024728768039494753, + 0.0035626108292490244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.08903924375772476 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486505150794983 + ], + "opposite_logprob_tokens": [ + -1.2236504554748535 + ], + "logprob_reward": -0.3486505150794983, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486505150794983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13746215403079987, + 0.11928346008062363, + 0.24404087662696838, + 0.11560613662004471, + 0.04697732627391815, + 0.06791723519563675, + 0.16552656888961792, + 0.033197417855262756, + 0.05085992440581322, + 0.019128920510411263 + ], + "credit_map": [ + 0.014060256071388721, + 0.020083637908101082, + 0.02153884992003441, + 0.014755912125110626, + 0.01546296663582325, + 0.011505335569381714, + 0.019155461341142654, + 0.020899731665849686, + 0.014667795971035957, + 0.006792465690523386, + 0.041015639901161194, + 0.01802404411137104, + 0.005848660599440336, + 0.007098822388797998, + 0.009305092506110668, + 0.01653093844652176, + 0.022287921980023384, + 0.010819912888109684, + 0.00952980201691389, + 0.008337163366377354, + 0.014076258055865765, + 0.08458664268255234, + 0.01275639422237873, + 0.08164678514003754, + 0.015306244604289532, + 0.014292989857494831, + 0.016467800363898277, + 0.013059667311608791, + 0.01216314360499382, + 0.012236627750098705, + 0.018448881804943085, + 0.01363077387213707, + 0.004349166993051767, + 0.006002867594361305, + 0.006818929687142372, + 0.00865339208394289, + 0.00555399339646101, + 0.004109312314540148, + 0.005482375621795654, + 0.0060072909109294415, + 0.009773117490112782, + 0.004622426349669695, + 0.011234791949391365, + 0.005188229493796825, + 0.004962977021932602, + 0.006858726032078266, + 0.008426045067608356, + 0.016850922256708145, + 0.01433776505291462, + 0.011633871123194695, + 0.04061608761548996, + 0.016058294102549553, + 0.01604810543358326, + 0.01541109848767519, + 0.023067258298397064, + 0.028354080393910408, + 0.005369268823415041, + 0.0025300588458776474, + 0.005059107206761837, + 0.004213572014123201, + 0.0025886737275868654, + 0.004855257458984852, + 0.0037587350234389305, + 0.004822747316211462, + 0.0051583279855549335, + 0.003988875076174736, + 0.011905967257916927, + 0.0048709530383348465, + 0.004129896406084299, + 0.006097511854022741, + 0.005247355904430151, + 0.00946103222668171, + 0.00224804668687284, + 0.001988715725019574, + 0.0030456066597253084, + 0.002409646986052394, + 0.0018091435777023435, + 0.0021874047815799713, + 0.0020879150833934546, + 0.003352442057803273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.3486505150794983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520989775657654 + ], + "opposite_logprob_tokens": [ + -1.5020990371704102 + ], + "logprob_reward": -0.2520989775657654, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520989775657654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08909855037927628, + 0.20307858288288116, + 0.24818365275859833, + 0.2642105519771576, + 0.028826259076595306, + 0.02202121540904045, + 0.034530214965343475, + 0.04785747081041336, + 0.038641151040792465, + 0.023552313446998596 + ], + "credit_map": [ + 0.008305584080517292, + 0.014180994592607021, + 0.011462094262242317, + 0.008606579154729843, + 0.012955931946635246, + 0.007696179207414389, + 0.01126580685377121, + 0.014625382609665394, + 0.020881420001387596, + 0.011755922809243202, + 0.06940916925668716, + 0.020634792745113373, + 0.00823051668703556, + 0.012418528087437153, + 0.013764847069978714, + 0.045983392745256424, + 0.024522805586457253, + 0.015015787445008755, + 0.010323894210159779, + 0.009170815348625183, + 0.01855141669511795, + 0.07484100759029388, + 0.0174225103110075, + 0.07833540439605713, + 0.028200676664710045, + 0.021454453468322754, + 0.07721243798732758, + 0.029627161100506783, + 0.021953687071800232, + 0.027186298742890358, + 0.028523912653326988, + 0.030051928013563156, + 0.0030925741884857416, + 0.003549060318619013, + 0.0034514283761382103, + 0.004111820366233587, + 0.00354686938226223, + 0.0031002468895167112, + 0.0036562387831509113, + 0.004318022169172764, + 0.002570006065070629, + 0.0021793649066239595, + 0.004265698604285717, + 0.0026675101835280657, + 0.00254278932698071, + 0.0022232255432754755, + 0.0028728751931339502, + 0.0026997465174645185, + 0.0038508486468344927, + 0.003478906350210309, + 0.006243844050914049, + 0.0030110336374491453, + 0.0028965508099645376, + 0.00427013635635376, + 0.00534448865801096, + 0.005434403661638498, + 0.004859535954892635, + 0.003533654846251011, + 0.006513187661767006, + 0.0075033013708889484, + 0.005626363679766655, + 0.005536814220249653, + 0.009460086934268475, + 0.004824523814022541, + 0.004388056229799986, + 0.0042665088549256325, + 0.008157145231962204, + 0.0042306799441576, + 0.004305095877498388, + 0.0029631699435412884, + 0.004795308690518141, + 0.005535186268389225, + 0.0022380035370588303, + 0.0024069761857390404, + 0.00376874883659184, + 0.0033713229931890965, + 0.0030796772334724665, + 0.00256728264503181, + 0.0028751983772963285, + 0.003245103871449828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.2520989775657654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow, majestic, wide-angle, wraparound shot, we gaze from a godlike perspective upon a vast protoplanetary disk, engaged in a magnificent gyroscope, dominating the entire expanse of the horizon. This vast cosmic vortex, composed of countless clouds of deep purple and fiery orange gas and wispy dust lanes, rotates steadily and leisurely within the blackness of space. Pulled inexorably by the powerful gravity of a central, immature protostar, the flow of matter spirals inexorably inward, forming vast gravitational vortices and denser clumps of matter. In this process, countless dust grains collide and coalesce, emitting bursts of nascent light, like countless flickering embers scattered across the velvety curtain of space. These faint, warm flashes illuminate the dark regions within the disk, heralding the birth and formation of a new world from chaos. The background of the entire picture is the deep interstellar space, dotted with distant and cold stars. The overall color tone transitions between the warm incandescent light in the center and the cold blue tones at the outer edge, creating a tranquil atmosphere that is both magnificent and full of hope for life, perfectly fitting a humanistic documentary's poetic exploration of the origin of the universe.", + "seed": 29736, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_8/index_1377_seed_29736_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": -0.1894852936267853, + "vlm_logprob": -0.1894852936267853, + "vlm_yes_no_margin": 2.149999976158142, + "vlm_token_credit": -0.1894852936267853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020754076540470123 + ], + "opposite_logprob_tokens": [ + -4.145753860473633 + ], + "logprob_reward": -0.020754076540470123, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020754076540470123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22500242292881012, + 0.04325711354613304, + 0.03923691809177399, + 0.019590135663747787, + 0.1009773388504982, + 0.17336100339889526, + 0.1514619141817093, + 0.03759089112281799, + 0.11853427439928055, + 0.09098793566226959 + ], + "credit_map": [ + 0.017224064096808434, + 0.00970140565186739, + 0.07658212631940842, + 0.017438022419810295, + 0.010770930908620358, + 0.06844712048768997, + 0.011515231803059578, + 0.013323538936674595, + 0.0054704975336790085, + 0.0052267261780798435, + 0.006382952444255352, + 0.008166885934770107, + 0.00498252734541893, + 0.003153599100187421, + 0.006388104986399412, + 0.0034858197905123234, + 0.0035769662354141474, + 0.0038620533887296915, + 0.007030965760350227, + 0.00968486163765192, + 0.0038285301998257637, + 0.0030617353040724993, + 0.00463994313031435, + 0.0035518682561814785, + 0.002010217634961009, + 0.0027275856118649244, + 0.0026869350112974644, + 0.0022681127302348614, + 0.0025159840006381273, + 0.002133216243237257, + 0.002863850211724639, + 0.0023842351511120796, + 0.007924018427729607, + 0.005562543403357267, + 0.011397075839340687, + 0.011468877084553242, + 0.0065359012223780155, + 0.011292130686342716, + 0.021598942577838898, + 0.025197863578796387, + 0.03254435211420059, + 0.015414568595588207, + 0.0342162661254406, + 0.021154873073101044, + 0.009936805814504623, + 0.013710894621908665, + 0.019444484263658524, + 0.026938775554299355, + 0.01522460300475359, + 0.04736035317182541, + 0.011718395166099072, + 0.011034256778657436, + 0.014853435568511486, + 0.009553351439535618, + 0.012033644132316113, + 0.029683886095881462, + 0.004462830256670713, + 0.005519503727555275, + 0.00854549277573824, + 0.0043646832928061485, + 0.0028282159473747015, + 0.002985167782753706, + 0.004073088988661766, + 0.004811910912394524, + 0.0210590697824955, + 0.009819523431360722, + 0.017237650230526924, + 0.013589505106210709, + 0.011462961323559284, + 0.01512930728495121, + 0.013349704444408417, + 0.016886569559574127, + 0.013084791600704193, + 0.00953923724591732, + 0.009596985764801502, + 0.01274095568805933, + 0.010814455337822437, + 0.011316444724798203, + 0.014099916443228722, + 0.009795156307518482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.020754076540470123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905079424381256 + ], + "opposite_logprob_tokens": [ + -2.5790507793426514 + ], + "logprob_reward": -0.07905079424381256, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905079424381256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18315696716308594, + 0.0695597380399704, + 0.05037504434585571, + 0.04070017859339714, + 0.18424127995967865, + 0.06704721599817276, + 0.05172223970293999, + 0.21590982377529144, + 0.09064475446939468, + 0.04664278030395508 + ], + "credit_map": [ + 0.012127745896577835, + 0.006255457643419504, + 0.065468430519104, + 0.010322253219783306, + 0.007352821063250303, + 0.06250365078449249, + 0.010929591953754425, + 0.008197029121220112, + 0.00523749366402626, + 0.0069670164957642555, + 0.013911046087741852, + 0.006563140545040369, + 0.008380915969610214, + 0.004812425933778286, + 0.01799163408577442, + 0.0056960745714604855, + 0.004102107137441635, + 0.005483979359269142, + 0.009683948941528797, + 0.012137264013290405, + 0.004279905464500189, + 0.004236223641782999, + 0.005709990859031677, + 0.0047416286543011665, + 0.005088575650006533, + 0.005541314836591482, + 0.004998234566301107, + 0.004878359381109476, + 0.004905425012111664, + 0.005614710506051779, + 0.005027944687753916, + 0.004645616747438908, + 0.033366184681653976, + 0.02124176174402237, + 0.02837478369474411, + 0.017695214599370956, + 0.01210272591561079, + 0.02624468319118023, + 0.01644679345190525, + 0.028769144788384438, + 0.009106761775910854, + 0.008196358568966389, + 0.009078138507902622, + 0.008174004964530468, + 0.007679152302443981, + 0.006579842418432236, + 0.006649706978350878, + 0.011583262123167515, + 0.005083936266601086, + 0.009069913998246193, + 0.005904352758079767, + 0.005293898284435272, + 0.005736472550779581, + 0.0038838887121528387, + 0.006504065822809935, + 0.010245716199278831, + 0.016861794516444206, + 0.07005723565816879, + 0.01463955920189619, + 0.01348594855517149, + 0.01736685074865818, + 0.010847033932805061, + 0.01839422434568405, + 0.054257191717624664, + 0.015279632061719894, + 0.007668640464544296, + 0.01326832640916109, + 0.010631554760038853, + 0.009109883569180965, + 0.0114572299644351, + 0.010638568550348282, + 0.012590927071869373, + 0.007467153016477823, + 0.004930973052978516, + 0.00347497477196157, + 0.006769279483705759, + 0.006097353994846344, + 0.004793177824467421, + 0.007252479903399944, + 0.005857394076883793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.07905079424381256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003215700387955 + ], + "opposite_logprob_tokens": [ + -2.6950321197509766 + ], + "logprob_reward": -0.07003215700387955, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003215700387955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062265802174806595, + 0.024693563580513, + 0.03347446769475937, + 0.023306604474782944, + 0.09324327111244202, + 0.2897880971431732, + 0.048079714179039, + 0.352176696062088, + 0.05082184821367264, + 0.022149967029690742 + ], + "credit_map": [ + 0.005222540348768234, + 0.004132621455937624, + 0.01768679916858673, + 0.006475662812590599, + 0.004592655226588249, + 0.014445233158767223, + 0.004310449119657278, + 0.005399842746555805, + 0.003580732038244605, + 0.0025609598960727453, + 0.0032192831858992577, + 0.004221636336296797, + 0.002476072171702981, + 0.0023605702444911003, + 0.0037987541873008013, + 0.002475557615980506, + 0.0042913067154586315, + 0.0027731554582715034, + 0.005294895265251398, + 0.005712269805371761, + 0.0033643355127424, + 0.0033137553837150335, + 0.006235632114112377, + 0.0024891197681427, + 0.002968631684780121, + 0.0024212633725255728, + 0.0050233411602675915, + 0.0031833902467042208, + 0.00218098983168602, + 0.002280962886288762, + 0.0022519235499203205, + 0.002996101975440979, + 0.01669093780219555, + 0.009390433318912983, + 0.013801740482449532, + 0.009248289279639721, + 0.005810359492897987, + 0.012639530934393406, + 0.00928040686994791, + 0.016381576657295227, + 0.05364203825592995, + 0.024883965030312538, + 0.058562982827425, + 0.03866573050618172, + 0.01598312333226204, + 0.020866742357611656, + 0.03013378009200096, + 0.04704974964261055, + 0.005244647618383169, + 0.01546978484839201, + 0.003742135362699628, + 0.003724933136254549, + 0.005400827154517174, + 0.0037643704563379288, + 0.0038939272053539753, + 0.006839090026915073, + 0.02831749990582466, + 0.11176853626966476, + 0.027088334783911705, + 0.02053552120923996, + 0.02466653473675251, + 0.018617697060108185, + 0.03274286538362503, + 0.0884397029876709, + 0.009067893028259277, + 0.004173358902335167, + 0.00735008530318737, + 0.005536401644349098, + 0.004224939737468958, + 0.007373845670372248, + 0.005716286599636078, + 0.007379040587693453, + 0.0034012366086244583, + 0.0020324510987848043, + 0.0022571885492652655, + 0.003463270142674446, + 0.0021109289955347776, + 0.002139379968866706, + 0.003453576937317848, + 0.0032919354271143675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.07003215700387955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893892526626587 + ], + "opposite_logprob_tokens": [ + -1.053938865661621 + ], + "logprob_reward": -0.42893892526626587, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42893892526626587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06406106054782867, + 0.020642872899770737, + 0.018296224996447563, + 0.032367728650569916, + 0.03325756639242172, + 0.346561998128891, + 0.06871004402637482, + 0.2672050893306732, + 0.12740594148635864, + 0.021491453051567078 + ], + "credit_map": [ + 0.003993872087448835, + 0.002126691397279501, + 0.02323749288916588, + 0.004714347887784243, + 0.0023653623647987843, + 0.02066728100180626, + 0.0034403575118631124, + 0.0035156516823917627, + 0.002492372179403901, + 0.0027232279535382986, + 0.0027761554811149836, + 0.004502239637076855, + 0.0023678154684603214, + 0.0014074245700612664, + 0.002880541840568185, + 0.0014930958859622478, + 0.0019188926089555025, + 0.001735943602398038, + 0.0031587230041623116, + 0.0038206670433282852, + 0.0016432999400421977, + 0.0018557492876425385, + 0.0025394398253411055, + 0.0016235093353316188, + 0.003558433847501874, + 0.0038784502539783716, + 0.004015838261693716, + 0.0039705620147287846, + 0.003991617821156979, + 0.004206126555800438, + 0.0051246145740151405, + 0.0036220846232026815, + 0.001798491575755179, + 0.0023972608614712954, + 0.0036510296631604433, + 0.004428657237440348, + 0.0024738784413784742, + 0.0032136377412825823, + 0.008456463925540447, + 0.006838146131485701, + 0.048699814826250076, + 0.07575233280658722, + 0.0445377491414547, + 0.03476318344473839, + 0.023628350347280502, + 0.03993627056479454, + 0.025524374097585678, + 0.05371993035078049, + 0.006080973893404007, + 0.01972099393606186, + 0.005592386703938246, + 0.004030909389257431, + 0.006392933893948793, + 0.0038931502494961023, + 0.005625234451144934, + 0.0173734612762928, + 0.0220437403768301, + 0.08809441328048706, + 0.01595289073884487, + 0.014382152818143368, + 0.0192180834710598, + 0.014307115226984024, + 0.02460884302854538, + 0.06859787553548813, + 0.022529544308781624, + 0.010070274583995342, + 0.017201445996761322, + 0.01686227321624756, + 0.010590133257210255, + 0.01731286756694317, + 0.014207400381565094, + 0.018632005900144577, + 0.004348027054220438, + 0.00199856492690742, + 0.0020300575997680426, + 0.0027008617762476206, + 0.001912964042276144, + 0.001969554927200079, + 0.003486897796392441, + 0.0030445249285548925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.42893892526626587 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486505150794983 + ], + "opposite_logprob_tokens": [ + -1.2236504554748535 + ], + "logprob_reward": -0.3486505150794983, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486505150794983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14582234621047974, + 0.05633700639009476, + 0.0534709133207798, + 0.06412138789892197, + 0.1393943727016449, + 0.07039406895637512, + 0.09420423209667206, + 0.15773996710777283, + 0.09607281535863876, + 0.12244284898042679 + ], + "credit_map": [ + 0.011110788211226463, + 0.007312897127121687, + 0.047249685972929, + 0.012904842384159565, + 0.008922906592488289, + 0.040661998093128204, + 0.008596397005021572, + 0.009062845259904861, + 0.007561022881418467, + 0.006385671440511942, + 0.008126083761453629, + 0.009149676188826561, + 0.006063569337129593, + 0.0041746520437300205, + 0.00996912270784378, + 0.004907202906906605, + 0.00875035859644413, + 0.004851784091442823, + 0.00951085053384304, + 0.008650819770991802, + 0.005377782508730888, + 0.004993433132767677, + 0.00712585961446166, + 0.004210025072097778, + 0.009376497939229012, + 0.00835162028670311, + 0.008180588483810425, + 0.0088039580732584, + 0.006941787898540497, + 0.006710071116685867, + 0.008281701244413853, + 0.0074751609936356544, + 0.02558530494570732, + 0.01348479837179184, + 0.02057238109409809, + 0.01151491329073906, + 0.010888515040278435, + 0.01873224601149559, + 0.010416822507977486, + 0.028199397027492523, + 0.010082556866109371, + 0.015138008631765842, + 0.009384376928210258, + 0.007213891483843327, + 0.0054442849941551685, + 0.006815293803811073, + 0.007079312112182379, + 0.009236341342329979, + 0.010715439915657043, + 0.02594813145697117, + 0.009724547155201435, + 0.008106363005936146, + 0.00968171562999487, + 0.007762524764984846, + 0.00913254451006651, + 0.013132967054843903, + 0.011959882453083992, + 0.044654134660959244, + 0.016581406816840172, + 0.011863313615322113, + 0.010852187871932983, + 0.009599607437849045, + 0.016479773446917534, + 0.035749658942222595, + 0.015602570958435535, + 0.00920114479959011, + 0.016467714682221413, + 0.01097646914422512, + 0.010518192313611507, + 0.009868024848401546, + 0.010704530403017998, + 0.01273417193442583, + 0.022357406094670296, + 0.010819531977176666, + 0.008074026554822922, + 0.01761823520064354, + 0.011830795556306839, + 0.012338655069470406, + 0.021827585995197296, + 0.017576608806848526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.3486505150794983 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slow, majestic, wide-angle, wraparound shot, we gaze from a godlike perspective upon a vast protoplanetary disk, engaged in a magnificent gyroscope, dominating the entire expanse of the horizon. This vast cosmic vortex, composed of countless clouds of deep purple and fiery orange gas and wispy dust lanes, rotates steadily and leisurely within the blackness of space. Pulled inexorably by the powerful gravity of a central, immature protostar, the flow of matter spirals inexorably inward, forming vast gravitational vortices and denser clumps of matter. In this process, countless dust grains collide and coalesce, emitting bursts of nascent light, like countless flickering embers scattered across the velvety curtain of space. These faint, warm flashes illuminate the dark regions within the disk, heralding the birth and formation of a new world from chaos. The background of the entire picture is the deep interstellar space, dotted with distant and cold stars. The overall color tone transitions between the warm incandescent light in the center and the cold blue tones at the outer edge, creating a tranquil atmosphere that is both magnificent and full of hope for life, perfectly fitting a humanistic documentary's poetic exploration of the origin of the universe.", + "seed": 29737, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_8/index_1377_seed_29737_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": -0.23083948567509652, + "vlm_logprob": -0.23083948567509652, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.23083948567509652, + "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 central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0383894182741642 + ], + "opposite_logprob_tokens": [ + -3.4133894443511963 + ], + "logprob_reward": -0.0383894182741642, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0383894182741642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07815977931022644, + 0.49032673239707947, + 0.02208447828888893, + 0.014403150416910648, + 0.013204331509768963, + 0.04726610332727432, + 0.0810912549495697, + 0.1840788722038269, + 0.03784431144595146, + 0.031541045755147934 + ], + "credit_map": [ + 0.008757071569561958, + 0.0070952740497887135, + 0.012396316975355148, + 0.010608099400997162, + 0.011100517585873604, + 0.011124169453978539, + 0.009771153330802917, + 0.007307174615561962, + 0.02188935875892639, + 0.17195959389209747, + 0.021863603964447975, + 0.017059888690710068, + 0.023646434769034386, + 0.19658221304416656, + 0.013258806429803371, + 0.024066805839538574, + 0.004959617275744677, + 0.001726868562400341, + 0.002428942359983921, + 0.002097496995702386, + 0.0012750349706038833, + 0.002621517051011324, + 0.002329027745872736, + 0.004645972978323698, + 0.0018745206762105227, + 0.0015946015482768416, + 0.0017348771216347814, + 0.0015647595282644033, + 0.0011399139184504747, + 0.0018616386223584414, + 0.0017796071479097009, + 0.0028532310388982296, + 0.002043709624558687, + 0.0016050600679591298, + 0.0015273004537448287, + 0.001418190193362534, + 0.0012849343474954367, + 0.0014478903030976653, + 0.0015670740976929665, + 0.002310172189027071, + 0.004038728307932615, + 0.011899684555828571, + 0.012792506255209446, + 0.0032660258002579212, + 0.0022604549303650856, + 0.003893673187121749, + 0.004126050043851137, + 0.004988974425941706, + 0.004801442380994558, + 0.029070526361465454, + 0.0033675036393105984, + 0.0036876697558909655, + 0.003921190742403269, + 0.02598913572728634, + 0.005015027709305286, + 0.005238752346485853, + 0.02894987165927887, + 0.014469592832028866, + 0.04907602816820145, + 0.014221751131117344, + 0.010770587250590324, + 0.028325561434030533, + 0.016269078478217125, + 0.02199639566242695, + 0.005029597785323858, + 0.00412737438455224, + 0.005851993337273598, + 0.004560454748570919, + 0.0035056695342063904, + 0.004142809193581343, + 0.0052302535623312, + 0.005396154243499041, + 0.004706425126641989, + 0.0023776108864694834, + 0.004973210394382477, + 0.0051610576920211315, + 0.003936832770705223, + 0.0027523525059223175, + 0.003936134744435549, + 0.003697420237585902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.0383894182741642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12714999914169312 + ], + "opposite_logprob_tokens": [ + -2.127150058746338 + ], + "logprob_reward": -0.12714999914169312, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12714999914169312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053261496126651764, + 0.09682127088308334, + 0.015227532014250755, + 0.0192538071423769, + 0.024524196982383728, + 0.06537947058677673, + 0.06512381881475449, + 0.3395472466945648, + 0.24400092661380768, + 0.07686026394367218 + ], + "credit_map": [ + 0.008302176371216774, + 0.005457589868456125, + 0.006545678712427616, + 0.004926683381199837, + 0.006559643894433975, + 0.011148534715175629, + 0.005172825884073973, + 0.005148367490619421, + 0.0049550398252904415, + 0.03218383342027664, + 0.004628869704902172, + 0.004092401824891567, + 0.00478648766875267, + 0.03749355673789978, + 0.003603715216740966, + 0.0050773643888533115, + 0.0026626994367688894, + 0.0015390568878501654, + 0.0021399743854999542, + 0.0017630868824198842, + 0.0016520855715498328, + 0.0015520047163590789, + 0.0014341402566060424, + 0.0024844827130436897, + 0.002703872974961996, + 0.0038807366508990526, + 0.0023775079753249884, + 0.0015790060861036181, + 0.0014396858168765903, + 0.003183919470757246, + 0.0016835978021845222, + 0.0024054795503616333, + 0.003006512764841318, + 0.0033068887423723936, + 0.0026263392064720392, + 0.003215858479961753, + 0.0033321285154670477, + 0.002462047152221203, + 0.0029785670340061188, + 0.003595853690057993, + 0.003456712467595935, + 0.014829359948635101, + 0.024257401004433632, + 0.003987977281212807, + 0.0030964170582592487, + 0.007350658066570759, + 0.004360354505479336, + 0.004040595609694719, + 0.006071696523576975, + 0.022916343063116074, + 0.004429846070706844, + 0.0038102297112345695, + 0.0035053750034421682, + 0.016530780121684074, + 0.003356298664584756, + 0.004503248259425163, + 0.03830159083008766, + 0.022637875750660896, + 0.06160687655210495, + 0.04149891436100006, + 0.025663509964942932, + 0.047838445752859116, + 0.06501829624176025, + 0.036981742829084396, + 0.059940848499536514, + 0.021465418860316277, + 0.044588759541511536, + 0.023112555965781212, + 0.015539372339844704, + 0.026738567277789116, + 0.019586442038416862, + 0.03302894905209541, + 0.012008998543024063, + 0.007499211933463812, + 0.01493803784251213, + 0.007654430344700813, + 0.007788117974996567, + 0.008764003403484821, + 0.007932759821414948, + 0.010274697095155716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.12714999914169312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043075501918792725 + ], + "opposite_logprob_tokens": [ + -3.1680755615234375 + ], + "logprob_reward": -0.043075501918792725, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043075501918792725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03948044404387474, + 0.03747308999300003, + 0.01163884624838829, + 0.019090399146080017, + 0.014988991431891918, + 0.029859889298677444, + 0.07531079649925232, + 0.7325300574302673, + 0.013797983527183533, + 0.025829503312706947 + ], + "credit_map": [ + 0.005637587979435921, + 0.003715947736054659, + 0.004576926119625568, + 0.0033868218306452036, + 0.007105269003659487, + 0.00771469809114933, + 0.0037889177910983562, + 0.0035542785190045834, + 0.004878387786448002, + 0.009534917771816254, + 0.0028780712746083736, + 0.0019023638451471925, + 0.0027670010458678007, + 0.010026856325566769, + 0.0018466596957296133, + 0.003638830501586199, + 0.002363402396440506, + 0.0022520858328789473, + 0.0012644254602491856, + 0.0011208740761503577, + 0.000859994615893811, + 0.0017572185024619102, + 0.0008712816052138805, + 0.0011495633516460657, + 0.003736390033736825, + 0.0017216033302247524, + 0.002273751888424158, + 0.0015637341421097517, + 0.0013218587264418602, + 0.0017591222422197461, + 0.0025602879468351603, + 0.004153651185333729, + 0.002399259014055133, + 0.0013836267171427608, + 0.002034333534538746, + 0.0020127238240092993, + 0.002108606044203043, + 0.0014449787558987737, + 0.0016422243788838387, + 0.001963239163160324, + 0.0025122095830738544, + 0.005487716291099787, + 0.0038767431396991014, + 0.0015520155429840088, + 0.00204439926892519, + 0.009686675854027271, + 0.0020631609950214624, + 0.0026369679253548384, + 0.005934255197644234, + 0.02028781734406948, + 0.004462421406060457, + 0.003609717357903719, + 0.004261606372892857, + 0.025571651756763458, + 0.004868838470429182, + 0.006314484868198633, + 0.10284357517957687, + 0.054289549589157104, + 0.16667966544628143, + 0.07161931693553925, + 0.04923788458108902, + 0.11969912797212601, + 0.08740978688001633, + 0.08075112104415894, + 0.0025408847723156214, + 0.0014608604833483696, + 0.001765278517268598, + 0.0013984774705022573, + 0.0015612796414643526, + 0.0016914638690650463, + 0.0017007229616865516, + 0.0016790159279480577, + 0.0037395742256194353, + 0.0023432632442563772, + 0.004122486803680658, + 0.003589483443647623, + 0.003281170269474387, + 0.0029421099461615086, + 0.0028846191707998514, + 0.002926797606050968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.043075501918792725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934453845024109 + ], + "opposite_logprob_tokens": [ + -0.6934453845024109 + ], + "logprob_reward": -0.6934453845024109, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934453845024109, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09588518738746643, + 0.5859073996543884, + 0.02643788419663906, + 0.03591563552618027, + 0.04937126114964485, + 0.03827979043126106, + 0.03939247131347656, + 0.08672284334897995, + 0.02668834663927555, + 0.015399090945720673 + ], + "credit_map": [ + 0.015571534633636475, + 0.00877155177295208, + 0.015074195340275764, + 0.013478311710059643, + 0.00953544769436121, + 0.014609615318477154, + 0.010050496086478233, + 0.008794042281806469, + 0.03225502744317055, + 0.1978561133146286, + 0.029720962047576904, + 0.021431278437376022, + 0.029663367196917534, + 0.2262798398733139, + 0.017403610050678253, + 0.03129728510975838, + 0.0058362288400530815, + 0.003622660879045725, + 0.0027462136931717396, + 0.0027832346968352795, + 0.0016102936351671815, + 0.002128126798197627, + 0.002290663542225957, + 0.005420461297035217, + 0.006206742487847805, + 0.004391995258629322, + 0.004317637998610735, + 0.004615257028490305, + 0.002474763197824359, + 0.003884177887812257, + 0.004190647508949041, + 0.005834413692355156, + 0.012617120519280434, + 0.004487918224185705, + 0.005464047659188509, + 0.003523406106978655, + 0.004370618145912886, + 0.004120309371501207, + 0.005767281167209148, + 0.009020558558404446, + 0.008074920624494553, + 0.006247006822377443, + 0.006351912394165993, + 0.0029400542844086885, + 0.0021493348758667707, + 0.0029802238568663597, + 0.0032268320210278034, + 0.006309504620730877, + 0.00432384992018342, + 0.007553736213594675, + 0.004333817400038242, + 0.004967538174241781, + 0.0038017944898456335, + 0.0044341376051306725, + 0.005724764429032803, + 0.004252833314239979, + 0.005314128007739782, + 0.007402842864394188, + 0.010310450568795204, + 0.012640715576708317, + 0.008822876028716564, + 0.007814205251634121, + 0.023267442360520363, + 0.011150184087455273, + 0.004691473674029112, + 0.0030972380191087723, + 0.003744745859876275, + 0.004580324981361628, + 0.0022892761044204235, + 0.0019834283739328384, + 0.0028133769519627094, + 0.003488483838737011, + 0.0016887513920664787, + 0.0016995348269119859, + 0.001649938989430666, + 0.001955439569428563, + 0.002574821701273322, + 0.0014627197524532676, + 0.002213588450103998, + 0.00215429556556046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.6934453845024109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central object a protoplanetary disk composed of deep purple and fiery orange gas with wispy dust lanes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a central immature protostar visible pulling matter inward to form gravitational vortices and denser clumps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are faint warm flashes resembling flickering embers visible within the disk, illuminating dark regions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background deep interstellar space dotted with distant cold stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video use a wide-angle wraparound shot from a high godlike perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25213712453842163 + ], + "opposite_logprob_tokens": [ + -1.5021371841430664 + ], + "logprob_reward": -0.25213712453842163, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213712453842163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12771013379096985, + 0.12083519250154495, + 0.03231608867645264, + 0.02467159368097782, + 0.01872403733432293, + 0.2607922852039337, + 0.12843501567840576, + 0.17298856377601624, + 0.08037088811397552, + 0.03315618261694908 + ], + "credit_map": [ + 0.016583982855081558, + 0.013157368637621403, + 0.019369490444660187, + 0.015673359856009483, + 0.020993318408727646, + 0.016903169453144073, + 0.012689714320003986, + 0.012339738197624683, + 0.008154029957950115, + 0.03640032559633255, + 0.007624286226928234, + 0.006753734312951565, + 0.00725061260163784, + 0.0403592623770237, + 0.00566952396184206, + 0.008623410016298294, + 0.004060449544340372, + 0.005762195214629173, + 0.004033094737678766, + 0.004771578125655651, + 0.0024327761493623257, + 0.0027349325828254223, + 0.003483404638245702, + 0.005037657916545868, + 0.0030203356873244047, + 0.003018351038917899, + 0.0031441899482160807, + 0.0031756835523992777, + 0.002357697347179055, + 0.003553851507604122, + 0.0027316855266690254, + 0.00366979674436152, + 0.001934414729475975, + 0.0020319712348282337, + 0.002688178326934576, + 0.0032515455968677998, + 0.0033695173915475607, + 0.0014545355224981904, + 0.0020677363499999046, + 0.0019261379493400455, + 0.031926847994327545, + 0.06535744667053223, + 0.05743724852800369, + 0.019505390897393227, + 0.011484104208648205, + 0.025639524683356285, + 0.022397981956601143, + 0.027043748646974564, + 0.009768727235496044, + 0.03722072020173073, + 0.01376281026750803, + 0.013920457102358341, + 0.007744223810732365, + 0.022735469043254852, + 0.013264779932796955, + 0.01001783274114132, + 0.027812344953417778, + 0.01311651524156332, + 0.04467610642313957, + 0.013630367815494537, + 0.010472780093550682, + 0.026635337620973587, + 0.01627214066684246, + 0.020372990518808365, + 0.006835962645709515, + 0.005786753259599209, + 0.017808133736252785, + 0.009588148444890976, + 0.006944417487829924, + 0.009081256575882435, + 0.014690389856696129, + 0.009635819122195244, + 0.0053238882683217525, + 0.003925398923456669, + 0.003438901389017701, + 0.004491722211241722, + 0.0035231756046414375, + 0.004013290163129568, + 0.00403923075646162, + 0.0044005755335092545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.25213712453842163 + } + ], + "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/0000110/group_9/index_tensor([8406, 8406])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_9/index_tensor([8406, 8406])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..44aa9dd0b1aa52ae120f4368ae3fca3575f30c92 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_9/index_tensor([8406, 8406])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium-long shot, shot from an overhead angle, reveals a mountain-like, weathered skull at its center. Illuminated by a harsh, blue-white light cast from directly above, the background a boundless, pure black void. The light carves deep, sharp shadows into the skull's jagged, bony surface, its vast, hollow eye sockets and nasal cavity looming deep in the darkness. With a low, grinding sound, as if from the depths of the earth, tiny cracks begin to spread like spiderwebs across the skull's forehead. Then, the entire massive structure begins a slow, inexorable inward collapse. Chunks of yellowed bone peel away from the main structure and fall, silently disintegrating in mid-air, transforming into dense clouds of gray dust that flicker briefly in the cold light, like ghostly nebulae. The entire collapse process is steady and deliberate, with a sense of epic fate. Finally, the entire skull is completely dissolved into a swirling dust vortex in a silent implosion. When the last glimmering dust particle dissipates into nothingness, the sole light source slowly goes out, and the entire picture finally fades smoothly into absolute, silent darkness.", + "seed": 7317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_9/index_8406_seed_7317_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": -0.20786639163270593, + "vlm_logprob": -0.20786639163270593, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.20786639163270593, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5335102677345276 + ], + "opposite_logprob_tokens": [ + -0.9085102677345276 + ], + "logprob_reward": -0.5335102677345276, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5335102677345276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04681491479277611, + 0.03755669295787811, + 0.0921141654253006, + 0.06785330921411514, + 0.05373011529445648, + 0.02808314561843872, + 0.025439919903874397, + 0.5089136362075806, + 0.09685377031564713, + 0.0426403284072876 + ], + "credit_map": [ + 0.006136554758995771, + 0.0023088075686246157, + 0.006753971800208092, + 0.004138512536883354, + 0.002299076411873102, + 0.008648968301713467, + 0.006618748884648085, + 0.009910278022289276, + 0.007984858937561512, + 0.0029364072252064943, + 0.006158244330435991, + 0.0026094098575413227, + 0.0017020003870129585, + 0.006925803609192371, + 0.0023234800901263952, + 0.006916485261172056, + 0.01646694354712963, + 0.005199824925512075, + 0.0190709438174963, + 0.007439408451318741, + 0.005929899401962757, + 0.01565765216946602, + 0.007945971563458443, + 0.014403515495359898, + 0.007921393029391766, + 0.006926998496055603, + 0.007152675185352564, + 0.007079044822603464, + 0.003620010567829013, + 0.012842291966080666, + 0.008697493933141232, + 0.013613400980830193, + 0.006928229238837957, + 0.003151805605739355, + 0.00856839306652546, + 0.0063547720201313496, + 0.0030932226218283176, + 0.008489136584103107, + 0.008043545298278332, + 0.00910100806504488, + 0.0036127299536019564, + 0.002336282515898347, + 0.00317722768522799, + 0.002523040631785989, + 0.001957235625013709, + 0.00535464845597744, + 0.0031008771620690823, + 0.006021103356033564, + 0.004048170521855354, + 0.002053661970421672, + 0.004424223210662603, + 0.001779310405254364, + 0.0017140049021691084, + 0.0046486202627420425, + 0.0021192943677306175, + 0.004652634263038635, + 0.10212075710296631, + 0.022928910329937935, + 0.0744251012802124, + 0.020263107493519783, + 0.03890637308359146, + 0.10846778750419617, + 0.024820813909173012, + 0.11698079109191895, + 0.016351744532585144, + 0.006407750770449638, + 0.015927407890558243, + 0.0065125045366585255, + 0.006457183975726366, + 0.018658393993973732, + 0.00705112935975194, + 0.019487649202346802, + 0.005710287485271692, + 0.005670933052897453, + 0.00666428217664361, + 0.004541685804724693, + 0.004084236454218626, + 0.005192383658140898, + 0.005392774008214474, + 0.005383744370192289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.5335102677345276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067605809308588505 + ], + "opposite_logprob_tokens": [ + -5.006760597229004 + ], + "logprob_reward": -0.0067605809308588505, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067605809308588505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16708505153656006, + 0.06726956367492676, + 0.12857407331466675, + 0.07339465618133545, + 0.059161920100450516, + 0.015879346057772636, + 0.04257744178175926, + 0.225424662232399, + 0.08180765807628632, + 0.13882558047771454 + ], + "credit_map": [ + 0.029745876789093018, + 0.010452179238200188, + 0.024261167272925377, + 0.013372036628425121, + 0.00855820719152689, + 0.030219363048672676, + 0.01599632389843464, + 0.03447991982102394, + 0.01209663413465023, + 0.00572150107473135, + 0.012061560526490211, + 0.004880404099822044, + 0.0037294032517820597, + 0.012049142271280289, + 0.004387362394481897, + 0.012343555688858032, + 0.02137615531682968, + 0.00932062603533268, + 0.024758901447057724, + 0.013105093501508236, + 0.010388731025159359, + 0.019625242799520493, + 0.012319047935307026, + 0.017680292949080467, + 0.009878503158688545, + 0.010355252772569656, + 0.008754967711865902, + 0.00650967238470912, + 0.003966124728322029, + 0.012716701254248619, + 0.006417994387447834, + 0.014795449562370777, + 0.0069345650263130665, + 0.004840649664402008, + 0.011212974786758423, + 0.007390407379716635, + 0.004441704601049423, + 0.00867681298404932, + 0.00785202719271183, + 0.007812785916030407, + 0.0019056196324527264, + 0.0014348168624565005, + 0.0027218565810471773, + 0.0014498336240649223, + 0.0012555863941088319, + 0.0026183121372014284, + 0.0016847819788381457, + 0.0028085384983569384, + 0.006507858633995056, + 0.0030679956544190645, + 0.007375523913651705, + 0.004073325078934431, + 0.0019223252311348915, + 0.007709046360105276, + 0.004412450361996889, + 0.0075089228339493275, + 0.03546906262636185, + 0.014318596571683884, + 0.03431421145796776, + 0.014523107558488846, + 0.02075049839913845, + 0.04078271612524986, + 0.024442734196782112, + 0.04082374647259712, + 0.010115275159478188, + 0.00840007234364748, + 0.011390036903321743, + 0.008189253509044647, + 0.006303333677351475, + 0.01310237031430006, + 0.010434955358505249, + 0.01387237198650837, + 0.0141283068805933, + 0.01378917507827282, + 0.01718776486814022, + 0.010823377408087254, + 0.009840035811066628, + 0.033311013132333755, + 0.01007447112351656, + 0.02967144548892975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0067605809308588505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28138360381126404 + ], + "opposite_logprob_tokens": [ + -1.4063836336135864 + ], + "logprob_reward": -0.28138360381126404, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28138360381126404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052821457386016846, + 0.07085537165403366, + 0.08581305295228958, + 0.044940538704395294, + 0.035934992134571075, + 0.022075504064559937, + 0.04411150887608528, + 0.5588122010231018, + 0.04235658794641495, + 0.042278822511434555 + ], + "credit_map": [ + 0.01043983455747366, + 0.0034356610849499702, + 0.00595062505453825, + 0.0032477860804647207, + 0.0024546263739466667, + 0.01006320770829916, + 0.005671503953635693, + 0.011558216996490955, + 0.015116828493773937, + 0.0044645764864981174, + 0.011226425878703594, + 0.0049033816903829575, + 0.003692940343171358, + 0.012656072154641151, + 0.004587013740092516, + 0.014208137057721615, + 0.014612840488553047, + 0.005230287089943886, + 0.017415130510926247, + 0.007848680019378662, + 0.006306569091975689, + 0.013916670344769955, + 0.007374382112175226, + 0.013108497485518456, + 0.006768253166228533, + 0.004261244088411331, + 0.009616763330996037, + 0.003699255408719182, + 0.0026013192255049944, + 0.0070496597327291965, + 0.004015788435935974, + 0.006928257644176483, + 0.008010944351553917, + 0.0019870188552886248, + 0.006191169377416372, + 0.002336632227525115, + 0.0011000677477568388, + 0.006214858964085579, + 0.0027109673246741295, + 0.007383335381746292, + 0.0026287224609404802, + 0.0013496199389919639, + 0.0024811788462102413, + 0.0019288952462375164, + 0.001398257096298039, + 0.004672414157539606, + 0.0020837888587266207, + 0.0055326297879219055, + 0.00894975196570158, + 0.002715411828830838, + 0.008204814977943897, + 0.0032835882157087326, + 0.0017680424498394132, + 0.008144712075591087, + 0.002669889247044921, + 0.008375297300517559, + 0.11107142269611359, + 0.024662254378199577, + 0.08313357084989548, + 0.022325310856103897, + 0.04214567691087723, + 0.11970987170934677, + 0.026772649958729744, + 0.12899142503738403, + 0.007651177700608969, + 0.0036666197702288628, + 0.008965871296823025, + 0.003656198736280203, + 0.003986817318946123, + 0.005186325404793024, + 0.0036165842320770025, + 0.00562699418514967, + 0.004964166786521673, + 0.0043748789466917515, + 0.007512376178056002, + 0.0036806140560656786, + 0.005261819809675217, + 0.006152855232357979, + 0.0037354580126702785, + 0.006596657447516918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.28138360381126404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082914546132088 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082914546132088, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082914546132088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046566423028707504, + 0.017179714515805244, + 0.044546160846948624, + 0.04861659184098244, + 0.019816502928733826, + 0.03600539267063141, + 0.050509799271821976, + 0.5030457377433777, + 0.17722319066524506, + 0.056490521878004074 + ], + "credit_map": [ + 0.008180004544556141, + 0.003084641182795167, + 0.007070408668369055, + 0.0032965217251330614, + 0.002939287107437849, + 0.0077462028712034225, + 0.004837587010115385, + 0.009411772713065147, + 0.002046906854957342, + 0.0013162705581635237, + 0.002612727228552103, + 0.0015636214520782232, + 0.0009447646443732083, + 0.0033706501126289368, + 0.0017174136592075229, + 0.0036073606461286545, + 0.007483236957341433, + 0.0024559537414461374, + 0.008439341560006142, + 0.0032055904157459736, + 0.0025127336848527193, + 0.00806858204305172, + 0.0047827851958572865, + 0.007597939111292362, + 0.006655867211520672, + 0.006743558216840029, + 0.006747209932655096, + 0.004335048142820597, + 0.0024861681740731, + 0.008307174779474735, + 0.00440059183165431, + 0.008940977044403553, + 0.0028615552000701427, + 0.0013995419722050428, + 0.002334041055291891, + 0.0016990170115604997, + 0.0012118866434320807, + 0.0038754986599087715, + 0.001902861869893968, + 0.004532100632786751, + 0.0038857280742377043, + 0.004247110802680254, + 0.00465412950143218, + 0.005071334075182676, + 0.003764168592169881, + 0.004586330149322748, + 0.005261946469545364, + 0.004534645471721888, + 0.004408125299960375, + 0.0037841745652258396, + 0.005802868865430355, + 0.003744169371202588, + 0.0026339285541325808, + 0.012586710043251514, + 0.00508410669863224, + 0.012465713545680046, + 0.09857957065105438, + 0.020853223279118538, + 0.07645425945520401, + 0.019200274720788002, + 0.03699454292654991, + 0.10943445563316345, + 0.024387964978814125, + 0.11714140325784683, + 0.02898811176419258, + 0.014245182275772095, + 0.029760854318737984, + 0.014515831135213375, + 0.01643991470336914, + 0.02709735557436943, + 0.014560390263795853, + 0.03161553293466568, + 0.006655882112681866, + 0.007929464802145958, + 0.007622069679200649, + 0.0061585973016917706, + 0.006022523622959852, + 0.008458226919174194, + 0.006723508238792419, + 0.006920256186276674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.016082914546132088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159459114074707 + ], + "opposite_logprob_tokens": [ + -1.701594591140747 + ], + "logprob_reward": -0.20159459114074707, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159459114074707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0547664575278759, + 0.10593465715646744, + 0.03894907981157303, + 0.04183514416217804, + 0.030090557411313057, + 0.02743471786379814, + 0.08919556438922882, + 0.17233259975910187, + 0.2618183493614197, + 0.1776428520679474 + ], + "credit_map": [ + 0.012036498636007309, + 0.002639498095959425, + 0.008898929692804813, + 0.0033302446827292442, + 0.0026249100919812918, + 0.009661178104579449, + 0.004532806109637022, + 0.011042389087378979, + 0.018557149916887283, + 0.007990118116140366, + 0.016488270834088326, + 0.007294908631592989, + 0.005183734465390444, + 0.020706025883555412, + 0.007615074981004, + 0.022099366411566734, + 0.005906142760068178, + 0.0028431026730686426, + 0.007083030417561531, + 0.005257694981992245, + 0.0033486487809568644, + 0.005402549169957638, + 0.004223948810249567, + 0.004883958958089352, + 0.005403484683483839, + 0.006988466251641512, + 0.006120418198406696, + 0.00385516625829041, + 0.002919179853051901, + 0.0060249557718634605, + 0.004661650862544775, + 0.005861818324774504, + 0.004272913560271263, + 0.0022749248892068863, + 0.003744343761354685, + 0.003619251772761345, + 0.001898153219372034, + 0.004798985552042723, + 0.004952938295900822, + 0.004529044963419437, + 0.003628309816122055, + 0.002636837773025036, + 0.004101854749023914, + 0.003444911213591695, + 0.0022587713319808245, + 0.004290992394089699, + 0.0030611467082053423, + 0.004011891316622496, + 0.011294089257717133, + 0.009040847420692444, + 0.01750495284795761, + 0.0061645652167499065, + 0.0064787548035383224, + 0.014880137518048286, + 0.009283377788960934, + 0.0145488316193223, + 0.028250625357031822, + 0.010084173642098904, + 0.02054198645055294, + 0.01294029038399458, + 0.012969431467354298, + 0.03744330629706383, + 0.011292887851595879, + 0.03880990296602249, + 0.02648957632482052, + 0.009261677041649818, + 0.027396593242883682, + 0.011785419657826424, + 0.007679050788283348, + 0.0804753452539444, + 0.013776001520454884, + 0.08495467156171799, + 0.01607389748096466, + 0.01406070962548256, + 0.03190929815173149, + 0.01408801507204771, + 0.015507298521697521, + 0.0352170392870903, + 0.015299019403755665, + 0.03548755496740341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.20159459114074707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium-long shot, shot from an overhead angle, reveals a mountain-like, weathered skull at its center. Illuminated by a harsh, blue-white light cast from directly above, the background a boundless, pure black void. The light carves deep, sharp shadows into the skull's jagged, bony surface, its vast, hollow eye sockets and nasal cavity looming deep in the darkness. With a low, grinding sound, as if from the depths of the earth, tiny cracks begin to spread like spiderwebs across the skull's forehead. Then, the entire massive structure begins a slow, inexorable inward collapse. Chunks of yellowed bone peel away from the main structure and fall, silently disintegrating in mid-air, transforming into dense clouds of gray dust that flicker briefly in the cold light, like ghostly nebulae. The entire collapse process is steady and deliberate, with a sense of epic fate. Finally, the entire skull is completely dissolved into a swirling dust vortex in a silent implosion. When the last glimmering dust particle dissipates into nothingness, the sole light source slowly goes out, and the entire picture finally fades smoothly into absolute, silent darkness.", + "seed": 7318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_9/index_8406_seed_7318_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": -0.14449820015579462, + "vlm_logprob": -0.14449820015579462, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.14449820015579462, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.435293048620224 + ], + "opposite_logprob_tokens": [ + -1.0602930784225464 + ], + "logprob_reward": -0.435293048620224, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.435293048620224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07132382690906525, + 0.017415009438991547, + 0.022353846579790115, + 0.02333025261759758, + 0.6070640683174133, + 0.06152200698852539, + 0.03263358026742935, + 0.014322854578495026, + 0.0439654104411602, + 0.10606908798217773 + ], + "credit_map": [ + 0.011837918311357498, + 0.006948157679289579, + 0.013639661483466625, + 0.006697771139442921, + 0.006467098370194435, + 0.013697093352675438, + 0.003800102975219488, + 0.008236022666096687, + 0.002747912425547838, + 0.0016012273263186216, + 0.003279426135122776, + 0.0014437646605074406, + 0.0018373116618022323, + 0.0032464589457958937, + 0.00097786751575768, + 0.0022810413502156734, + 0.004061535466462374, + 0.0025041073095053434, + 0.004147605039179325, + 0.0023405887186527252, + 0.002750822575762868, + 0.002719540148973465, + 0.0016102262306958437, + 0.0022194210905581713, + 0.004510398022830486, + 0.0015157561283558607, + 0.004658104851841927, + 0.0016301590949296951, + 0.0017772893188521266, + 0.004337136633694172, + 0.0014419218059629202, + 0.0034594859462231398, + 0.11930501461029053, + 0.046788521111011505, + 0.10806059837341309, + 0.03555484861135483, + 0.04419678449630737, + 0.14915618300437927, + 0.03568122535943985, + 0.06832090020179749, + 0.0060165454633533955, + 0.007392807863652706, + 0.011477580294013023, + 0.004954514093697071, + 0.005157823674380779, + 0.014877540059387684, + 0.0059133353643119335, + 0.005731857847422361, + 0.004025292582809925, + 0.0037916756700724363, + 0.004193788859993219, + 0.0029149572364985943, + 0.004151393659412861, + 0.003933293744921684, + 0.0057222903706133366, + 0.003900890937075019, + 0.001867951825261116, + 0.0026360817719250917, + 0.0015987353399395943, + 0.0018264355603605509, + 0.0017595015233382583, + 0.002069157315418124, + 0.0011812584707513452, + 0.0013837335864081979, + 0.004090809728950262, + 0.005560860503464937, + 0.005674975924193859, + 0.006002818234264851, + 0.0048647369258105755, + 0.005989625584334135, + 0.005193619057536125, + 0.006587965413928032, + 0.008414747193455696, + 0.01922021247446537, + 0.012311463244259357, + 0.014231187291443348, + 0.016213208436965942, + 0.012928931973874569, + 0.0127300675958395, + 0.010019274428486824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.435293048620224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046560680493712425 + ], + "opposite_logprob_tokens": [ + -5.379655838012695 + ], + "logprob_reward": -0.0046560680493712425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046560680493712425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08460263162851334, + 0.04316183552145958, + 0.042165227234363556, + 0.06787795573472977, + 0.4425790011882782, + 0.03540242090821266, + 0.05400429293513298, + 0.051039405167102814, + 0.103669174015522, + 0.07549809664487839 + ], + "credit_map": [ + 0.01868380978703499, + 0.006521811243146658, + 0.01591063290834427, + 0.005396919324994087, + 0.00602326262742281, + 0.019839739426970482, + 0.004305942915380001, + 0.007920514792203903, + 0.007152769714593887, + 0.0029767341911792755, + 0.006588829215615988, + 0.0043141478672623634, + 0.003448174800723791, + 0.009936793707311153, + 0.002776550827547908, + 0.0059678321704268456, + 0.00723837036639452, + 0.004334787372499704, + 0.007738250773400068, + 0.003503573825582862, + 0.0036884674336761236, + 0.009693002328276634, + 0.0023581325076520443, + 0.003610640997067094, + 0.008500856347382069, + 0.005195487756282091, + 0.0134124755859375, + 0.00412345165386796, + 0.0044637201353907585, + 0.021463634446263313, + 0.002815800718963146, + 0.007902531884610653, + 0.09380073100328445, + 0.033719323575496674, + 0.0780310109257698, + 0.02550225518643856, + 0.03260962665081024, + 0.10125422477722168, + 0.02729550190269947, + 0.05036631226539612, + 0.006294784136116505, + 0.005119612906128168, + 0.006668964866548777, + 0.003261746373027563, + 0.0024101745802909136, + 0.005997410509735346, + 0.0020997582469135523, + 0.003549968358129263, + 0.005838869605213404, + 0.007996844127774239, + 0.006384823005646467, + 0.004010278731584549, + 0.007056142203509808, + 0.007178429048508406, + 0.008135090582072735, + 0.00740381283685565, + 0.005358056165277958, + 0.005878360942006111, + 0.0046660685911774635, + 0.006100897677242756, + 0.007729186676442623, + 0.005985924508422613, + 0.005818304605782032, + 0.009502611123025417, + 0.007840555161237717, + 0.012907801195979118, + 0.011673595756292343, + 0.015757722780108452, + 0.016194269061088562, + 0.013448566198348999, + 0.011453425511717796, + 0.014393242076039314, + 0.005792582407593727, + 0.01447529997676611, + 0.00878052320331335, + 0.0070162927731871605, + 0.012496420182287693, + 0.009716386906802654, + 0.00928119383752346, + 0.007939400151371956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.0046560680493712425 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215715169906616 + ], + "opposite_logprob_tokens": [ + -1.502157211303711 + ], + "logprob_reward": -0.25215715169906616, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215715169906616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0618232898414135, + 0.04864731431007385, + 0.027697807177901268, + 0.016901424154639244, + 0.4906664490699768, + 0.034786250442266464, + 0.03797442466020584, + 0.06862051784992218, + 0.08057420700788498, + 0.13230831921100616 + ], + "credit_map": [ + 0.01346003171056509, + 0.004305825103074312, + 0.012059839442372322, + 0.003977223765105009, + 0.004279643297195435, + 0.013648228719830513, + 0.0027995072305202484, + 0.007292990572750568, + 0.010702943429350853, + 0.0033785151317715645, + 0.008695708587765694, + 0.004710475914180279, + 0.004160615149885416, + 0.007331129163503647, + 0.003123900154605508, + 0.006544027011841536, + 0.005822852253913879, + 0.002920836443081498, + 0.005055594723671675, + 0.002340654144063592, + 0.002088633831590414, + 0.005534692667424679, + 0.001470528426580131, + 0.0024640150368213654, + 0.0019672480411827564, + 0.001102329813875258, + 0.00322054885327816, + 0.0009392370120622218, + 0.0010585280833765864, + 0.005693165119737387, + 0.0006555315922014415, + 0.0022648365702480078, + 0.09609245508909225, + 0.03617993742227554, + 0.08786297589540482, + 0.027467763051390648, + 0.03557024523615837, + 0.12235906720161438, + 0.028622640296816826, + 0.05651139095425606, + 0.004708890803158283, + 0.0021365482825785875, + 0.008763342164456844, + 0.002727977465838194, + 0.0021623261272907257, + 0.008241238072514534, + 0.00252927141264081, + 0.0035166582092642784, + 0.005039615090936422, + 0.004726724233478308, + 0.005593593697994947, + 0.0030655022710561752, + 0.003913831431418657, + 0.00597467552870512, + 0.006210125982761383, + 0.0034503552597016096, + 0.007254166528582573, + 0.011633021757006645, + 0.009189576841890812, + 0.009040631353855133, + 0.007401188835501671, + 0.008066453970968723, + 0.008219457231462002, + 0.007816021330654621, + 0.005987857468426228, + 0.008860942907631397, + 0.00663509638980031, + 0.009989491663873196, + 0.015002420172095299, + 0.011510303243994713, + 0.01193010900169611, + 0.010657988488674164, + 0.01103575062006712, + 0.025511693209409714, + 0.015176867134869099, + 0.012830398976802826, + 0.020435858517885208, + 0.016001272946596146, + 0.016640113666653633, + 0.014676361344754696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.25215715169906616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792031720280647 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009792031720280647, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792031720280647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03773292899131775, + 0.024084730073809624, + 0.02008110284805298, + 0.05252533778548241, + 0.40865859389305115, + 0.0990837812423706, + 0.08195199072360992, + 0.10527611523866653, + 0.07775750756263733, + 0.09284792840480804 + ], + "credit_map": [ + 0.00779427494853735, + 0.002707160310819745, + 0.008634469471871853, + 0.003314083442091942, + 0.0026870507281273603, + 0.006811941042542458, + 0.0021399324759840965, + 0.0036440175026655197, + 0.0031824889592826366, + 0.0015493830433115363, + 0.0047859144397079945, + 0.0013794316910207272, + 0.0018412548815831542, + 0.006692752707749605, + 0.0012973967241123319, + 0.0033561072777956724, + 0.0033641124609857798, + 0.0014855905901640654, + 0.003623240627348423, + 0.0019185889977961779, + 0.001674789935350418, + 0.004891628865152597, + 0.0011225627968087792, + 0.0020005893893539906, + 0.005763390101492405, + 0.003232745686545968, + 0.010705186985433102, + 0.0026552174240350723, + 0.002983094658702612, + 0.018298538401722908, + 0.0019199119415134192, + 0.006967256776988506, + 0.08260363340377808, + 0.03114033304154873, + 0.07184174656867981, + 0.024247057735919952, + 0.02987661398947239, + 0.09789992868900299, + 0.02516229823231697, + 0.04588698595762253, + 0.011976074427366257, + 0.00950644537806511, + 0.019763214513659477, + 0.008994827046990395, + 0.007892338559031487, + 0.01948615536093712, + 0.00882528256624937, + 0.012639452703297138, + 0.009765829890966415, + 0.008707424625754356, + 0.011038994416594505, + 0.006043202709406614, + 0.008734222501516342, + 0.012537834234535694, + 0.015581254847347736, + 0.009543228894472122, + 0.009915387257933617, + 0.018027666956186295, + 0.011264852248132229, + 0.012713435105979443, + 0.01570877991616726, + 0.012090219184756279, + 0.011902924627065659, + 0.01365285087376833, + 0.005564689636230469, + 0.008622458204627037, + 0.00558100501075387, + 0.008534573018550873, + 0.016558872535824776, + 0.011800503358244896, + 0.010733703151345253, + 0.010361704044044018, + 0.00699969008564949, + 0.016758274286985397, + 0.0112645598128438, + 0.009815587662160397, + 0.013780166395008564, + 0.011518768966197968, + 0.012673812918365002, + 0.010037064552307129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.009792031720280647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059270069003105 + ], + "opposite_logprob_tokens": [ + -3.89559268951416 + ], + "logprob_reward": -0.02059270069003105, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059270069003105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0752355083823204, + 0.04378605633974075, + 0.031223881989717484, + 0.024019712582230568, + 0.1393490731716156, + 0.04711588844656944, + 0.10717325657606125, + 0.08766086399555206, + 0.20045602321624756, + 0.243979811668396 + ], + "credit_map": [ + 0.013777317479252815, + 0.0063776555471122265, + 0.016168853268027306, + 0.006559652741998434, + 0.004951419308781624, + 0.015203947201371193, + 0.004194688051939011, + 0.008001967333257198, + 0.007220639381557703, + 0.00263741216622293, + 0.009097893722355366, + 0.002745807869359851, + 0.0033851033076643944, + 0.011548216454684734, + 0.0020310627296566963, + 0.005119917448610067, + 0.0059854392893612385, + 0.002881689229980111, + 0.005620266310870647, + 0.003517443547025323, + 0.0034947581589221954, + 0.00502134021371603, + 0.001978389685973525, + 0.0027245525270700455, + 0.003927041310817003, + 0.0021629270631819963, + 0.004660492762923241, + 0.0014000077499076724, + 0.0017666453495621681, + 0.006346866488456726, + 0.0012993313139304519, + 0.0024563989136368036, + 0.015830038115382195, + 0.009247363545000553, + 0.028658416122198105, + 0.006731580011546612, + 0.009346960112452507, + 0.04696383327245712, + 0.00679801357910037, + 0.0157728660851717, + 0.006595550570636988, + 0.006866895593702793, + 0.007995415478944778, + 0.004551329649984837, + 0.0039042236749082804, + 0.009596213698387146, + 0.0030599001329392195, + 0.004546354990452528, + 0.016811756417155266, + 0.012852099724113941, + 0.013697538524866104, + 0.009558591991662979, + 0.011246620677411556, + 0.01531197875738144, + 0.017750412225723267, + 0.009944248013198376, + 0.00818649586290121, + 0.01674131490290165, + 0.009923514910042286, + 0.008909326046705246, + 0.014953948557376862, + 0.01156727783381939, + 0.008844578638672829, + 0.008534400723874569, + 0.017952559515833855, + 0.02405940741300583, + 0.01801767572760582, + 0.020104166120290756, + 0.037775278091430664, + 0.0284084752202034, + 0.024728180840611458, + 0.029410263523459435, + 0.018123110756278038, + 0.03558271750807762, + 0.02686227299273014, + 0.0379197932779789, + 0.042395371943712234, + 0.030094115063548088, + 0.030264830216765404, + 0.022737577557563782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.02059270069003105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium-long shot, shot from an overhead angle, reveals a mountain-like, weathered skull at its center. Illuminated by a harsh, blue-white light cast from directly above, the background a boundless, pure black void. The light carves deep, sharp shadows into the skull's jagged, bony surface, its vast, hollow eye sockets and nasal cavity looming deep in the darkness. With a low, grinding sound, as if from the depths of the earth, tiny cracks begin to spread like spiderwebs across the skull's forehead. Then, the entire massive structure begins a slow, inexorable inward collapse. Chunks of yellowed bone peel away from the main structure and fall, silently disintegrating in mid-air, transforming into dense clouds of gray dust that flicker briefly in the cold light, like ghostly nebulae. The entire collapse process is steady and deliberate, with a sense of epic fate. Finally, the entire skull is completely dissolved into a swirling dust vortex in a silent implosion. When the last glimmering dust particle dissipates into nothingness, the sole light source slowly goes out, and the entire picture finally fades smoothly into absolute, silent darkness.", + "seed": 7319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_9/index_8406_seed_7319_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": -0.23587299697101116, + "vlm_logprob": -0.23587299697101116, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.23587299697101116, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8317687511444092 + ], + "opposite_logprob_tokens": [ + -0.5817687511444092 + ], + "logprob_reward": -0.8317687511444092, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8317687511444092, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07379462569952011, + 0.05102669447660446, + 0.08774358034133911, + 0.10051456838846207, + 0.17291037738323212, + 0.06905804574489594, + 0.15444496273994446, + 0.06756994873285294, + 0.1557939052581787, + 0.0671432614326477 + ], + "credit_map": [ + 0.008279873058199883, + 0.005379310809075832, + 0.006884847301989794, + 0.0054946658201515675, + 0.011805910617113113, + 0.013181714341044426, + 0.0072074332274496555, + 0.01556087564677, + 0.007890435867011547, + 0.006004779599606991, + 0.004647926427423954, + 0.00644514337182045, + 0.006003675051033497, + 0.006634368561208248, + 0.005050376523286104, + 0.008349986746907234, + 0.01305996160954237, + 0.006797914393246174, + 0.00893419049680233, + 0.007126583717763424, + 0.010334434919059277, + 0.014368962496519089, + 0.009111235849559307, + 0.018010299652814865, + 0.012324599549174309, + 0.008174527436494827, + 0.011493992991745472, + 0.005258327350020409, + 0.00948034506291151, + 0.022061802446842194, + 0.007686510682106018, + 0.024034468457102776, + 0.010742012411355972, + 0.010236809961497784, + 0.008344228379428387, + 0.0077933985739946365, + 0.013551340438425541, + 0.05776949226856232, + 0.010867328383028507, + 0.05360577255487442, + 0.008342740125954151, + 0.006861947476863861, + 0.006193763576447964, + 0.008211188018321991, + 0.010981653816998005, + 0.008317260071635246, + 0.008160968311131, + 0.01198852900415659, + 0.01895962841808796, + 0.018046092242002487, + 0.014132899232208729, + 0.015945959836244583, + 0.020827509462833405, + 0.02494855970144272, + 0.011995754204690456, + 0.029588554054498672, + 0.005720058921724558, + 0.0045629628002643585, + 0.004543912131339312, + 0.004408818669617176, + 0.007217311765998602, + 0.014236239716410637, + 0.006371511146426201, + 0.020509131252765656, + 0.014529572799801826, + 0.018305707722902298, + 0.016467582434415817, + 0.01664884015917778, + 0.017639707773923874, + 0.024265803396701813, + 0.016127146780490875, + 0.031809546053409576, + 0.006935387384146452, + 0.007787309121340513, + 0.0073648556135594845, + 0.00861375406384468, + 0.00765429949387908, + 0.00879721250385046, + 0.009439856745302677, + 0.010550589300692081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.8317687511444092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014222213998436928 + ], + "opposite_logprob_tokens": [ + -4.264222145080566 + ], + "logprob_reward": -0.014222213998436928, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014222213998436928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07828173786401749, + 0.08560065925121307, + 0.03714956343173981, + 0.024961205199360847, + 0.3498353958129883, + 0.029886573553085327, + 0.07434093952178955, + 0.024029703810811043, + 0.28043684363365173, + 0.015477355569601059 + ], + "credit_map": [ + 0.007030858192592859, + 0.0069132898934185505, + 0.007662732619792223, + 0.0067251394502818584, + 0.009011921472847462, + 0.0163488257676363, + 0.006721022073179483, + 0.017867952585220337, + 0.01249760203063488, + 0.009193849749863148, + 0.006930881645530462, + 0.007204256020486355, + 0.007174143102020025, + 0.01616601273417473, + 0.006226147525012493, + 0.02020777016878128, + 0.0050451247952878475, + 0.0024967764038592577, + 0.005108812358230352, + 0.002521457150578499, + 0.003921103663742542, + 0.006525970995426178, + 0.004279828164726496, + 0.007250492926687002, + 0.0031121240463107824, + 0.001667879056185484, + 0.0031798428390175104, + 0.0018524267943575978, + 0.0019364027539268136, + 0.0072316695004701614, + 0.0020689356606453657, + 0.003911928739398718, + 0.013810144737362862, + 0.00885589886456728, + 0.019048476591706276, + 0.005899548530578613, + 0.011176449246704578, + 0.15401776134967804, + 0.010781468823552132, + 0.12624569237232208, + 0.0031734625808894634, + 0.0029020027723163366, + 0.002885161666199565, + 0.0034052326809614897, + 0.00358229735866189, + 0.004521740134805441, + 0.0026978417299687862, + 0.006718837656080723, + 0.006472983863204718, + 0.011081483215093613, + 0.004657086916267872, + 0.005739408079534769, + 0.017048144713044167, + 0.010623584501445293, + 0.006263075862079859, + 0.012455174699425697, + 0.0018559559248387814, + 0.0023456199560314417, + 0.0013541225343942642, + 0.0014370528515428305, + 0.0032205216120928526, + 0.004468159284442663, + 0.0028980430215597153, + 0.006450230255723, + 0.031574733555316925, + 0.04496138170361519, + 0.028391724452376366, + 0.024889789521694183, + 0.049445513635873795, + 0.04575152322649956, + 0.027260424569249153, + 0.02816176787018776, + 0.0012634712038561702, + 0.0018279188079759479, + 0.002554615493863821, + 0.0026097758673131466, + 0.002098439261317253, + 0.0018270887667313218, + 0.0014843667158856988, + 0.001811680500395596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.014222213998436928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709088623523712 + ], + "opposite_logprob_tokens": [ + -2.1270909309387207 + ], + "logprob_reward": -0.12709088623523712, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709088623523712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05747078359127045, + 0.018384015187621117, + 0.05902764946222305, + 0.029236767441034317, + 0.36695438623428345, + 0.08310046792030334, + 0.24550530314445496, + 0.05567645654082298, + 0.06345786154270172, + 0.02118629589676857 + ], + "credit_map": [ + 0.009248850867152214, + 0.004841690883040428, + 0.006757011637091637, + 0.004614064935594797, + 0.009527451358735561, + 0.008491182699799538, + 0.004990447778254747, + 0.009000087156891823, + 0.003358061658218503, + 0.0014700294705107808, + 0.002609577728435397, + 0.001213172683492303, + 0.00172779755666852, + 0.002872948069125414, + 0.001870961976237595, + 0.0032614690717309713, + 0.00940027367323637, + 0.0038389519322663546, + 0.006908264011144638, + 0.003945563919842243, + 0.006434169132262468, + 0.009508322924375534, + 0.005550054367631674, + 0.013442049734294415, + 0.003850777866318822, + 0.0020500689279288054, + 0.003085779957473278, + 0.001813554554246366, + 0.0032515409402549267, + 0.0064588794484734535, + 0.0021431343629956245, + 0.006583032663911581, + 0.015011361800134182, + 0.012575485743582249, + 0.020105835050344467, + 0.010125193744897842, + 0.01237486582249403, + 0.15703335404396057, + 0.014054740779101849, + 0.1256735473871231, + 0.007751686032861471, + 0.005843352992087603, + 0.01081728283315897, + 0.0064035505056381226, + 0.012018341571092606, + 0.012057687155902386, + 0.007859977893531322, + 0.020348595455288887, + 0.01941649056971073, + 0.029067177325487137, + 0.01742243394255638, + 0.0197026077657938, + 0.04566045477986336, + 0.039345063269138336, + 0.020051660016179085, + 0.054839424788951874, + 0.004585550632327795, + 0.005868778098374605, + 0.003421938978135586, + 0.0034974142909049988, + 0.00937303900718689, + 0.009746246039867401, + 0.00599701190367341, + 0.013186481781303883, + 0.003817520570009947, + 0.004561638459563255, + 0.0041043744422495365, + 0.006187338847666979, + 0.007304121740162373, + 0.011525684967637062, + 0.008720609359443188, + 0.017236577346920967, + 0.0016852733679115772, + 0.002374056028202176, + 0.0026543154381215572, + 0.0031295185908675194, + 0.0024818892125040293, + 0.0029907466378062963, + 0.0028406998608261347, + 0.0030297976918518543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.12709088623523712 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263621024787426 + ], + "opposite_logprob_tokens": [ + -3.651362180709839 + ], + "logprob_reward": -0.0263621024787426, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263621024787426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02734186314046383, + 0.060081589967012405, + 0.04537695273756981, + 0.02977614663541317, + 0.45996642112731934, + 0.03447315841913223, + 0.10733242332935333, + 0.034366756677627563, + 0.14745143055915833, + 0.05383330211043358 + ], + "credit_map": [ + 0.0029304535128176212, + 0.00279232463799417, + 0.0028508147224783897, + 0.0029260951559990644, + 0.005795842036604881, + 0.0032336211297661066, + 0.0031566445250064135, + 0.0036560643929988146, + 0.010168909095227718, + 0.005559816025197506, + 0.004954418167471886, + 0.0038585136644542217, + 0.004825049079954624, + 0.011665779165923595, + 0.004274488892406225, + 0.014774610288441181, + 0.006357711739838123, + 0.003085807431489229, + 0.006067432463169098, + 0.0032381010241806507, + 0.005112424958497286, + 0.007493289653211832, + 0.0052614035084843636, + 0.008760778233408928, + 0.00375151913613081, + 0.00223258207552135, + 0.003614104352891445, + 0.001968483440577984, + 0.0027730122674256563, + 0.007997028529644012, + 0.0024550005327910185, + 0.004984414204955101, + 0.018499918282032013, + 0.012386872433125973, + 0.024014616385102272, + 0.007117996457964182, + 0.01563788577914238, + 0.20125257968902588, + 0.014265653677284718, + 0.1667909026145935, + 0.0029864127282053232, + 0.002350070048123598, + 0.004545687232166529, + 0.002845947863534093, + 0.004041669890284538, + 0.00557360053062439, + 0.0026124659925699234, + 0.009517299942672253, + 0.011572455987334251, + 0.013382306322455406, + 0.008466779254376888, + 0.009284637868404388, + 0.01848115585744381, + 0.016804613173007965, + 0.00805180799216032, + 0.021288657560944557, + 0.0031503888312727213, + 0.004540279041975737, + 0.0019519012421369553, + 0.002094398019835353, + 0.005264102481305599, + 0.005838626530021429, + 0.003909112419933081, + 0.007617942523211241, + 0.015475344844162464, + 0.019381413236260414, + 0.01474829576909542, + 0.015662364661693573, + 0.023647813126444817, + 0.02331085316836834, + 0.016837848350405693, + 0.018387487158179283, + 0.003921728581190109, + 0.00664925342425704, + 0.006745447404682636, + 0.008348561823368073, + 0.007377743721008301, + 0.006811563391238451, + 0.006054679863154888, + 0.007924314588308334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.0263621024787426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992103099822998 + ], + "opposite_logprob_tokens": [ + -1.80492103099823 + ], + "logprob_reward": -0.17992103099822998, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992103099822998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032309580594301224, + 0.043498653918504715, + 0.029538914561271667, + 0.02591054141521454, + 0.35384687781333923, + 0.024465341120958328, + 0.20564110577106476, + 0.06152012571692467, + 0.07089334726333618, + 0.15237551927566528 + ], + "credit_map": [ + 0.004014281090348959, + 0.003381666261702776, + 0.003278204472735524, + 0.0031047917436808348, + 0.005915722344070673, + 0.004311832599341869, + 0.0030160099267959595, + 0.005287069361656904, + 0.007365087978541851, + 0.0038342922925949097, + 0.003688667668029666, + 0.0023440395016223192, + 0.003964613191783428, + 0.008192656561732292, + 0.0033775977790355682, + 0.01073169894516468, + 0.004078465048223734, + 0.002386230742558837, + 0.004417021758854389, + 0.0018573935376480222, + 0.0034177808556705713, + 0.005421425215899944, + 0.0031579930800944567, + 0.004802603740245104, + 0.0037780501879751682, + 0.0019341005245223641, + 0.002808369230479002, + 0.0013920515775680542, + 0.0022611739113926888, + 0.005822467152029276, + 0.0019118408672511578, + 0.006002487149089575, + 0.012264086864888668, + 0.009983119554817677, + 0.019250579178333282, + 0.007666586432605982, + 0.010288180783390999, + 0.1557672619819641, + 0.011599626392126083, + 0.12702742218971252, + 0.003124429378658533, + 0.0020153524819761515, + 0.0029041299130767584, + 0.0019229487515985966, + 0.0033873349893838167, + 0.003947846591472626, + 0.001906567718833685, + 0.00525673246011138, + 0.021658947691321373, + 0.021605292335152626, + 0.016928521916270256, + 0.01893441192805767, + 0.02922837622463703, + 0.03425908088684082, + 0.016815178096294403, + 0.04621129855513573, + 0.005973943509161472, + 0.008307915180921555, + 0.003850333159789443, + 0.004450127482414246, + 0.011653302237391472, + 0.009426340460777283, + 0.006273387465626001, + 0.011584771797060966, + 0.006556010339409113, + 0.006939120125025511, + 0.006378091871738434, + 0.008224017918109894, + 0.009455899707973003, + 0.01097854133695364, + 0.010471082292497158, + 0.011890583671629429, + 0.012534922920167446, + 0.018074484542012215, + 0.02082144096493721, + 0.02511841058731079, + 0.02213711105287075, + 0.018686119467020035, + 0.014816190116107464, + 0.020186850801110268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.17992103099822998 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium-long shot, shot from an overhead angle, reveals a mountain-like, weathered skull at its center. Illuminated by a harsh, blue-white light cast from directly above, the background a boundless, pure black void. The light carves deep, sharp shadows into the skull's jagged, bony surface, its vast, hollow eye sockets and nasal cavity looming deep in the darkness. With a low, grinding sound, as if from the depths of the earth, tiny cracks begin to spread like spiderwebs across the skull's forehead. Then, the entire massive structure begins a slow, inexorable inward collapse. Chunks of yellowed bone peel away from the main structure and fall, silently disintegrating in mid-air, transforming into dense clouds of gray dust that flicker briefly in the cold light, like ghostly nebulae. The entire collapse process is steady and deliberate, with a sense of epic fate. Finally, the entire skull is completely dissolved into a swirling dust vortex in a silent implosion. When the last glimmering dust particle dissipates into nothingness, the sole light source slowly goes out, and the entire picture finally fades smoothly into absolute, silent darkness.", + "seed": 7320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000110/group_9/index_8406_seed_7320_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": -0.152884969022125, + "vlm_logprob": -0.152884969022125, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.152884969022125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3928057849407196 + ], + "opposite_logprob_tokens": [ + -1.142805814743042 + ], + "logprob_reward": -0.3928057849407196, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3928057849407196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22754129767417908, + 0.08051493018865585, + 0.0781276598572731, + 0.07669933885335922, + 0.03475973382592201, + 0.09331932663917542, + 0.07510189712047577, + 0.10238421708345413, + 0.17262932658195496, + 0.058922287076711655 + ], + "credit_map": [ + 0.04870130121707916, + 0.011643975041806698, + 0.029950890690088272, + 0.00894925557076931, + 0.015185490250587463, + 0.06934604048728943, + 0.010706593282520771, + 0.03305774927139282, + 0.013385734520852566, + 0.0038171305786818266, + 0.009959944523870945, + 0.003950952086597681, + 0.005686004180461168, + 0.024305982515215874, + 0.0056815664283931255, + 0.013727615587413311, + 0.013981680385768414, + 0.005916454363614321, + 0.010755309835076332, + 0.004824104253202677, + 0.004618022590875626, + 0.017820890992879868, + 0.006889618467539549, + 0.013321581296622753, + 0.011593841016292572, + 0.008095446974039078, + 0.010835458524525166, + 0.008172008208930492, + 0.004688942339271307, + 0.01466861367225647, + 0.00691933324560523, + 0.011725696735084057, + 0.007046843413263559, + 0.002390914596617222, + 0.0065628038719296455, + 0.0030314945615828037, + 0.002612503245472908, + 0.005332109518349171, + 0.0038099007215350866, + 0.0039731659926474094, + 0.01981704868376255, + 0.005057026632130146, + 0.03074798360466957, + 0.006371039431542158, + 0.004409021232277155, + 0.006953938864171505, + 0.006142311729490757, + 0.013820948079228401, + 0.010837922804057598, + 0.008463393896818161, + 0.014612688682973385, + 0.006417024414986372, + 0.007472915109246969, + 0.009790908545255661, + 0.00748229306191206, + 0.010024753399193287, + 0.021364176645874977, + 0.009773138910531998, + 0.016848353669047356, + 0.009934138506650925, + 0.011479830369353294, + 0.009416820481419563, + 0.0069144549779593945, + 0.01665329746901989, + 0.04356946796178818, + 0.018635034561157227, + 0.023616736754775047, + 0.015469806268811226, + 0.011512220837175846, + 0.018894759938120842, + 0.011902238242328167, + 0.029029060155153275, + 0.006507277954369783, + 0.006537828128784895, + 0.01061651948839426, + 0.007088656537234783, + 0.007465054281055927, + 0.007281765341758728, + 0.006588314194232225, + 0.006836870685219765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.3928057849407196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650125306099653 + ], + "opposite_logprob_tokens": [ + -4.882649898529053 + ], + "logprob_reward": -0.007650125306099653, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650125306099653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11709555983543396, + 0.034110307693481445, + 0.020240530371665955, + 0.01851092278957367, + 0.014429863542318344, + 0.013265213929116726, + 0.027896961197257042, + 0.03756586089730263, + 0.660860538482666, + 0.05602419748902321 + ], + "credit_map": [ + 0.02201894484460354, + 0.006639572326093912, + 0.013996892608702183, + 0.005071498919278383, + 0.008601168170571327, + 0.03585805743932724, + 0.00796082429587841, + 0.016948610544204712, + 0.005616279318928719, + 0.002744576195254922, + 0.004471711348742247, + 0.0023561264388263226, + 0.003279963508248329, + 0.008376291953027248, + 0.002848659409210086, + 0.004416698589920998, + 0.0037114149890840054, + 0.002021105494350195, + 0.003227733541280031, + 0.0020458928775042295, + 0.0018561349716037512, + 0.003299583215266466, + 0.0017925722058862448, + 0.002286092611029744, + 0.00371809140779078, + 0.0016566269332543015, + 0.0031448560766875744, + 0.0014249284286051989, + 0.0011526270536705852, + 0.002630589297041297, + 0.0019623616244643927, + 0.0028208426665514708, + 0.002616801066324115, + 0.0011231800308451056, + 0.002971155336126685, + 0.0013110358268022537, + 0.0008112370851449668, + 0.001465431647375226, + 0.0018323061522096395, + 0.0022987164556980133, + 0.0025795495603233576, + 0.0008449344313703477, + 0.002064756117761135, + 0.0009164629736915231, + 0.0009057230781763792, + 0.0014943300047889352, + 0.0014691113028675318, + 0.0029903454706072807, + 0.004371757619082928, + 0.0027018447872251272, + 0.004474367015063763, + 0.001814522547647357, + 0.0022718177642673254, + 0.004282931797206402, + 0.0018744058907032013, + 0.006105311214923859, + 0.005971526261419058, + 0.003954373765736818, + 0.006465816404670477, + 0.0030495645478367805, + 0.006151122972369194, + 0.004653531592339277, + 0.0032719257287681103, + 0.004047998692840338, + 0.1990475058555603, + 0.058078695088624954, + 0.09176880121231079, + 0.030277874320745468, + 0.042884863913059235, + 0.06029617041349411, + 0.028145689517259598, + 0.15036097168922424, + 0.005192236974835396, + 0.0067456006072461605, + 0.008102554827928543, + 0.008029036223888397, + 0.006391219794750214, + 0.007606275379657745, + 0.005709834396839142, + 0.0082474360242486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.007650125306099653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813730835914612 + ], + "opposite_logprob_tokens": [ + -1.4063730239868164 + ], + "logprob_reward": -0.2813730835914612, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813730835914612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24065423011779785, + 0.10977030545473099, + 0.03655793145298958, + 0.03000297024846077, + 0.03488834947347641, + 0.03976886346936226, + 0.038741178810596466, + 0.07574092596769333, + 0.35492637753486633, + 0.038948867470026016 + ], + "credit_map": [ + 0.04748234525322914, + 0.014715597964823246, + 0.03217754140496254, + 0.008284220471978188, + 0.017464982345700264, + 0.07322312146425247, + 0.011389948427677155, + 0.035916466265916824, + 0.019189713522791862, + 0.00955718569457531, + 0.0146924564614892, + 0.007111749611794949, + 0.007433616556227207, + 0.025784660130739212, + 0.00884957890957594, + 0.01715134084224701, + 0.007918663322925568, + 0.0029006025288254023, + 0.004141406621783972, + 0.0028466572985053062, + 0.0023423039820045233, + 0.008159230463206768, + 0.0024694812018424273, + 0.005779586266726255, + 0.005451352335512638, + 0.0030132890678942204, + 0.003969026263803244, + 0.0031622229143977165, + 0.002107806969434023, + 0.005037567112594843, + 0.003426781389862299, + 0.0038349241949617863, + 0.009873674251139164, + 0.002266763476654887, + 0.008587641641497612, + 0.002023450331762433, + 0.0020397850312292576, + 0.004239000380039215, + 0.0025735492818057537, + 0.0032844848465174437, + 0.011035875417292118, + 0.0026142511051148176, + 0.007499512284994125, + 0.002210476668551564, + 0.002135321730747819, + 0.005409375764429569, + 0.0024392958730459213, + 0.00642475625500083, + 0.006993974559009075, + 0.002669421723112464, + 0.011802874505519867, + 0.0020575555972754955, + 0.0025808352511376143, + 0.0046662455424666405, + 0.001686605392023921, + 0.006283664144575596, + 0.010538667440414429, + 0.015454317443072796, + 0.01062193512916565, + 0.00450320215895772, + 0.005413104314357042, + 0.008634806610643864, + 0.0025640795938670635, + 0.018010806292295456, + 0.08412570506334305, + 0.03745283931493759, + 0.046395737677812576, + 0.024267492815852165, + 0.0344681441783905, + 0.037209611386060715, + 0.018514376133680344, + 0.07249246537685394, + 0.003940539434552193, + 0.004572414327412844, + 0.005213042721152306, + 0.004091902170330286, + 0.005949593614786863, + 0.004251578822731972, + 0.005287933629006147, + 0.005641863215714693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.2813730835914612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551363557577133 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551363557577133, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551363557577133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057584792375564575, + 0.011921466328203678, + 0.00970084872096777, + 0.009316538460552692, + 0.010253812186419964, + 0.034054871648550034, + 0.04230688512325287, + 0.14152301847934723, + 0.649104654788971, + 0.03423316031694412 + ], + "credit_map": [ + 0.011891947127878666, + 0.003459468949586153, + 0.007111095357686281, + 0.0022395795676857233, + 0.004478024784475565, + 0.016890687867999077, + 0.003058350645005703, + 0.008455641567707062, + 0.0022818667348474264, + 0.0009769004536792636, + 0.0021208336111158133, + 0.0008651801035739481, + 0.0010661614360287786, + 0.002303278772160411, + 0.0009657490882091224, + 0.0013414963614195585, + 0.0018680780194699764, + 0.0008387507405132055, + 0.00175445806235075, + 0.0007696631364524364, + 0.0007279948331415653, + 0.001488315756432712, + 0.0009240963263437152, + 0.0013294924283400178, + 0.0019219431560486555, + 0.0007614303613081574, + 0.0017671792302280664, + 0.0007112633902579546, + 0.0006498977891169488, + 0.001301151467487216, + 0.00099678966216743, + 0.0012068836949765682, + 0.0016313893720507622, + 0.0006139811011962593, + 0.002294046338647604, + 0.0006436179392039776, + 0.000614446762483567, + 0.0015109312953427434, + 0.0008854026673361659, + 0.0020599965937435627, + 0.0074382382445037365, + 0.000989131978712976, + 0.01273912750184536, + 0.0014321320923045278, + 0.0010410051327198744, + 0.002642597770318389, + 0.0013484311057254672, + 0.006424206309020519, + 0.008613627403974533, + 0.0028707708697766066, + 0.011987573467195034, + 0.0022676452063024044, + 0.002219327026978135, + 0.005570592358708382, + 0.0016369280638173223, + 0.007140420377254486, + 0.0475098080933094, + 0.03088272549211979, + 0.015569379553198814, + 0.008182059973478317, + 0.012529009953141212, + 0.014103379100561142, + 0.0055176070891320705, + 0.00722903897985816, + 0.18614734709262848, + 0.06062901392579079, + 0.08446609973907471, + 0.03664766252040863, + 0.04223211854696274, + 0.05682936683297157, + 0.030489740893244743, + 0.15166325867176056, + 0.003995839972048998, + 0.0041163042187690735, + 0.004540132824331522, + 0.004458739422261715, + 0.004781331401318312, + 0.003283006837591529, + 0.004654898773878813, + 0.00440290616825223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.012551363557577133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium-long shot from an overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a mountain-like, weathered skull with jagged, bony surfaces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a boundless, pure black void illuminated by a harsh, blue-white light from directly above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do tiny cracks spread across the skull's forehead and does the structure begin a slow inward collapse with chunks of yellowed bone peeling away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the skull completely dissolve into a swirling dust vortex and fade smoothly into absolute darkness as the light source goes out?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004448771476746 + ], + "opposite_logprob_tokens": [ + -2.69504451751709 + ], + "logprob_reward": -0.07004448771476746, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004448771476746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04207981005311012, + 0.01628575474023819, + 0.015529846772551537, + 0.015938175842165947, + 0.021806277334690094, + 0.052389875054359436, + 0.02585279382765293, + 0.11874841898679733, + 0.5093821287155151, + 0.1819869726896286 + ], + "credit_map": [ + 0.008136258460581303, + 0.003090235637500882, + 0.007514323107898235, + 0.003008166328072548, + 0.0026371804997324944, + 0.008475612848997116, + 0.003014198038727045, + 0.006203831173479557, + 0.0025627731811255217, + 0.0016285321908071637, + 0.0018609227845445275, + 0.0016942780930548906, + 0.0011786648537963629, + 0.00201446283608675, + 0.0033588744699954987, + 0.0019872456323355436, + 0.0020791797433048487, + 0.0013536798069253564, + 0.001911223866045475, + 0.001456199330277741, + 0.001257032621651888, + 0.003664134768769145, + 0.0015449771890416741, + 0.002263416536152363, + 0.0035100902896374464, + 0.0013650866458192468, + 0.002041878178715706, + 0.0012839880073443055, + 0.0013859494356438518, + 0.002966583939269185, + 0.0013343938626348972, + 0.0020502032712101936, + 0.002982625737786293, + 0.0019710208289325237, + 0.0023226209450513124, + 0.002448023995384574, + 0.00207160203717649, + 0.0035359859466552734, + 0.00247595994733274, + 0.003998434636741877, + 0.009573376737535, + 0.0023821727372705936, + 0.016541652381420135, + 0.002726615872234106, + 0.002475456800311804, + 0.005372451152652502, + 0.003201852785423398, + 0.010116293095052242, + 0.0065956879407167435, + 0.0025814448017627, + 0.00266695162281394, + 0.0017231624806299806, + 0.0019208075245842338, + 0.002618212252855301, + 0.0022472164127975702, + 0.005499310325831175, + 0.0442836619913578, + 0.007219376042485237, + 0.012984814122319221, + 0.005306037608534098, + 0.006284200120717287, + 0.009926455095410347, + 0.0061922152526676655, + 0.026551641523838043, + 0.12780816853046417, + 0.051201093941926956, + 0.06439865380525589, + 0.040302108973264694, + 0.03807611018419266, + 0.04365001618862152, + 0.03220351040363312, + 0.11174245178699493, + 0.01863294094800949, + 0.019964078441262245, + 0.02206282690167427, + 0.019947556778788567, + 0.02279653586447239, + 0.024291759356856346, + 0.021369924768805504, + 0.03292132914066315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.07004448771476746 + } + ], + "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/0000131/group_0/index_tensor([5465, 5465])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_0/index_tensor([5465, 5465])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..08bbf99028bb82cb2d5518603a91f06e36964ccc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_0/index_tensor([5465, 5465])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic tracking shot, shot from a low angle, pans rapidly to the right, closely following a modern police car in classic black and white livery. Late in the night, amidst a torrential downpour, the car suddenly enters the frame from the left, speeding off to the right across a wide, tarmacked city street at breakneck speed. Its body, thoroughly drenched by the rain, its glossy surface reflecting everything around it. The red and blue police lights atop the car flash in a rapid, rhythmic pattern, casting a dazzling light across the slippery, mirror-like pavement and the facades of the towering buildings lining the street, creating a constantly chasing, shifting pattern of red and blue. The car's bright headlights pierce the rain, casting twin sharp beams of light, while its rapidly spinning wheels send up streams of shimmering water. In the background, the glow of the city's neon signs and streetlights blends into a patchwork of blurred reflections on the waterlogged road, contrasting with the scattered, warm yellow lights filtering through the windows of distant buildings. The main color tone of the entire picture is cold, dominated by the deep night and cold rain. Only the red and blue police lights and neon colors become jumping highlights, creating a strong sense of urgency, tense atmosphere and dynamic urban rainy night aesthetics.", + "seed": 19254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_0/index_5465_seed_19254_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": -0.030877336487174036, + "vlm_logprob": -0.030877336487174036, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.030877336487174036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04484184458851814 + ], + "opposite_logprob_tokens": [ + -3.294841766357422 + ], + "logprob_reward": -0.04484184458851814, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04484184458851814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18514558672904968, + 0.0722799301147461, + 0.10487810522317886, + 0.0741315707564354, + 0.04899415746331215, + 0.15480774641036987, + 0.06834075599908829, + 0.09167856723070145, + 0.08543767780065536, + 0.11430587619543076 + ], + "credit_map": [ + 0.017405716702342033, + 0.02216861955821514, + 0.04013242945075035, + 0.02701612003147602, + 0.015539819374680519, + 0.010982674546539783, + 0.034673310816287994, + 0.017226893454790115, + 0.009575487114489079, + 0.007075316738337278, + 0.012034643441438675, + 0.013458686880767345, + 0.005228184629231691, + 0.008101082406938076, + 0.007931259460747242, + 0.008875268511474133, + 0.022952314466238022, + 0.01058110035955906, + 0.014883521944284439, + 0.011650493368506432, + 0.004751044325530529, + 0.01337397750467062, + 0.011385847814381123, + 0.015299816615879536, + 0.010616116225719452, + 0.005691851489245892, + 0.011812462471425533, + 0.009432979859411716, + 0.01021910272538662, + 0.008738256990909576, + 0.008940653875470161, + 0.008680148050189018, + 0.003375357249751687, + 0.005494983401149511, + 0.013586584478616714, + 0.006631982047110796, + 0.004154194612056017, + 0.004506009630858898, + 0.005305079743266106, + 0.005939963739365339, + 0.009225898422300816, + 0.011455398052930832, + 0.01788065955042839, + 0.01709207333624363, + 0.015542423352599144, + 0.029998216778039932, + 0.020146444439888, + 0.0334666408598423, + 0.008826503530144691, + 0.010131497867405415, + 0.013632411137223244, + 0.011244000867009163, + 0.0053206817246973515, + 0.005158571992069483, + 0.007129759062081575, + 0.006897332146763802, + 0.010946358554065228, + 0.009739507921040058, + 0.012212412431836128, + 0.011936482042074203, + 0.008671337738633156, + 0.011950327083468437, + 0.012326513417065144, + 0.01389563363045454, + 0.010886509902775288, + 0.01078761089593172, + 0.012487047351896763, + 0.011696802452206612, + 0.010703349485993385, + 0.0056265778839588165, + 0.010702765546739101, + 0.01254701241850853, + 0.017487088218331337, + 0.009342745877802372, + 0.028171559795737267, + 0.012826411984860897, + 0.008463922888040543, + 0.01117851585149765, + 0.018203478306531906, + 0.00863215234130621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.04484184458851814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373095020651817 + ], + "opposite_logprob_tokens": [ + -3.408730983734131 + ], + "logprob_reward": -0.03373095020651817, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373095020651817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3024134337902069, + 0.07880537211894989, + 0.0725511908531189, + 0.04267677292227745, + 0.025438359007239342, + 0.08847708255052567, + 0.08491377532482147, + 0.11885807663202286, + 0.05758605897426605, + 0.1282799392938614 + ], + "credit_map": [ + 0.024242743849754333, + 0.015092356130480766, + 0.08662772178649902, + 0.05219106376171112, + 0.02487245202064514, + 0.011963966302573681, + 0.02377215586602688, + 0.06365096569061279, + 0.007295165676623583, + 0.007376224733889103, + 0.0104301692917943, + 0.012108782306313515, + 0.004286124836653471, + 0.01249767653644085, + 0.0075729768723249435, + 0.01723824255168438, + 0.012403746135532856, + 0.006296097766608, + 0.012393401004374027, + 0.006786218844354153, + 0.004331582225859165, + 0.008856919594109058, + 0.009653111919760704, + 0.011830109171569347, + 0.005453141871839762, + 0.004319784697145224, + 0.006833845283836126, + 0.004177923314273357, + 0.006615199614316225, + 0.0059355637058615685, + 0.004929605405777693, + 0.004411707166582346, + 0.0019503963412716985, + 0.003258539130911231, + 0.004016777966171503, + 0.003490537405014038, + 0.002762923017144203, + 0.00289289653301239, + 0.003858211450278759, + 0.0032080758828669786, + 0.005343358963727951, + 0.008734443224966526, + 0.009688008576631546, + 0.009021032601594925, + 0.0073991864919662476, + 0.02214873768389225, + 0.006858694367110729, + 0.019283615052700043, + 0.007886144332587719, + 0.010728747583925724, + 0.017240768298506737, + 0.01250847801566124, + 0.008030550554394722, + 0.006328732706606388, + 0.007907149381935596, + 0.014283196069300175, + 0.023338191211223602, + 0.018662363290786743, + 0.01399683766067028, + 0.010894199833273888, + 0.010215301997959614, + 0.012514960952103138, + 0.018886519595980644, + 0.010349693708121777, + 0.01106172427535057, + 0.010381394065916538, + 0.004686923231929541, + 0.005513808224350214, + 0.007712197490036488, + 0.003307390259578824, + 0.005927200894802809, + 0.008995418436825275, + 0.012530462816357613, + 0.010120888240635395, + 0.026211390271782875, + 0.013507351279258728, + 0.013159376569092274, + 0.01679031364619732, + 0.01605393923819065, + 0.019906209781765938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.03373095020651817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016090303659439087 + ], + "opposite_logprob_tokens": [ + -4.141090393066406 + ], + "logprob_reward": -0.016090303659439087, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016090303659439087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08414270728826523, + 0.07832290232181549, + 0.07605206966400146, + 0.05025757476687431, + 0.0203254371881485, + 0.09372573345899582, + 0.10231684893369675, + 0.36153456568717957, + 0.06587466597557068, + 0.06744741648435593 + ], + "credit_map": [ + 0.00933430902659893, + 0.009069202467799187, + 0.010327641852200031, + 0.0107648316770792, + 0.005786321125924587, + 0.005404173396527767, + 0.01323054451495409, + 0.02022569067776203, + 0.007621770724654198, + 0.009023209102451801, + 0.008618492633104324, + 0.015154436230659485, + 0.005310039967298508, + 0.012155573815107346, + 0.007058835122734308, + 0.01338055171072483, + 0.012883632443845272, + 0.005539971403777599, + 0.0070726522244513035, + 0.008867627009749413, + 0.003760822582989931, + 0.011777595616877079, + 0.008495436049997807, + 0.017654333263635635, + 0.006604441441595554, + 0.0039887032471597195, + 0.009680714458227158, + 0.004663561470806599, + 0.006114921998232603, + 0.006393985822796822, + 0.005785305984318256, + 0.007025941740721464, + 0.001771015697158873, + 0.003168184543028474, + 0.0024912613444030285, + 0.0030391779728233814, + 0.0019405428320169449, + 0.0019233028870075941, + 0.0030474758241325617, + 0.002944477368146181, + 0.004604870919138193, + 0.0071692923083901405, + 0.00914281141012907, + 0.010190851986408234, + 0.008217833936214447, + 0.01970664970576763, + 0.024856846779584885, + 0.009836592711508274, + 0.009331300854682922, + 0.015678465366363525, + 0.020217739045619965, + 0.01621079258620739, + 0.010788066312670708, + 0.008516404777765274, + 0.008139780722558498, + 0.013434305787086487, + 0.05937260761857033, + 0.037625715136528015, + 0.039398178458213806, + 0.028987443074584007, + 0.03022870607674122, + 0.05435304716229439, + 0.05528882518410683, + 0.056280095130205154, + 0.008514629676938057, + 0.009798655286431313, + 0.012650835327804089, + 0.008785913698375225, + 0.0070625729858875275, + 0.00462523614987731, + 0.006535958498716354, + 0.007900867611169815, + 0.007689351215958595, + 0.0070872060023248196, + 0.008164196275174618, + 0.012172551825642586, + 0.008892742916941643, + 0.007952164858579636, + 0.0077598365023732185, + 0.007729377597570419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.016090303659439087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076981201767921 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011076981201767921, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076981201767921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1238088384270668, + 0.0346616692841053, + 0.08655299246311188, + 0.1479773074388504, + 0.027637293562293053, + 0.11003724485635757, + 0.14557389914989471, + 0.07925273478031158, + 0.07964421063661575, + 0.16485381126403809 + ], + "credit_map": [ + 0.009642560966312885, + 0.009503360837697983, + 0.024207549169659615, + 0.02063225768506527, + 0.007442161440849304, + 0.012662614695727825, + 0.022687148302793503, + 0.017031187191605568, + 0.004480189178138971, + 0.003879608353599906, + 0.00464586541056633, + 0.006685054395347834, + 0.002261695684865117, + 0.003984647337347269, + 0.004111289978027344, + 0.004613317549228668, + 0.012902554124593735, + 0.007800220977514982, + 0.007882723584771156, + 0.00841644685715437, + 0.007105420809239149, + 0.01479206420481205, + 0.01519682165235281, + 0.012456740252673626, + 0.0111932884901762, + 0.009278643876314163, + 0.011557276360690594, + 0.012333760038018227, + 0.025644127279520035, + 0.028864212334156036, + 0.013621092773973942, + 0.0354849137365818, + 0.0022996377665549517, + 0.0037547412794083357, + 0.004517214372754097, + 0.0038315283600240946, + 0.0021770878229290247, + 0.004178176634013653, + 0.003621880430728197, + 0.00325702503323555, + 0.005590461660176516, + 0.006974498275667429, + 0.006946077104657888, + 0.007672344800084829, + 0.009871202521026134, + 0.034986983984708786, + 0.008471268229186535, + 0.02952440455555916, + 0.022235436365008354, + 0.02881680428981781, + 0.02455427125096321, + 0.025255147367715836, + 0.01055110152810812, + 0.011971878819167614, + 0.01024505216628313, + 0.01194420363754034, + 0.014346998184919357, + 0.011125623248517513, + 0.009640970267355442, + 0.009633656591176987, + 0.00843129213899374, + 0.007080690935254097, + 0.012550233863294125, + 0.0064432667568326, + 0.009702327661216259, + 0.011554987169802189, + 0.015382654964923859, + 0.010191992856562138, + 0.009154071100056171, + 0.005087698344141245, + 0.008338036015629768, + 0.010232441127300262, + 0.01767292059957981, + 0.007066769991070032, + 0.055915843695402145, + 0.014170704409480095, + 0.008428332395851612, + 0.015566772781312466, + 0.0355401448905468, + 0.010492315515875816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.011076981201767921 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646602779626846 + ], + "opposite_logprob_tokens": [ + -3.0486466884613037 + ], + "logprob_reward": -0.048646602779626846, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646602779626846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10480345785617828, + 0.06558670103549957, + 0.04698450490832329, + 0.036613915115594864, + 0.05881889909505844, + 0.15187712013721466, + 0.12629543244838715, + 0.19858232140541077, + 0.06168536841869354, + 0.14875225722789764 + ], + "credit_map": [ + 0.007187315262854099, + 0.011893057264387608, + 0.023693367838859558, + 0.014898351393640041, + 0.009903245605528355, + 0.007261442020535469, + 0.010049802251160145, + 0.019916875287890434, + 0.006154166534543037, + 0.005572541616857052, + 0.00917566753923893, + 0.00949308555573225, + 0.0038042340893298388, + 0.00887953583151102, + 0.008879879489541054, + 0.013627585954964161, + 0.008096064440906048, + 0.005477388389408588, + 0.005585723556578159, + 0.0058656674809753895, + 0.0030086738988757133, + 0.006308962590992451, + 0.006485339254140854, + 0.006156681105494499, + 0.0038067386485636234, + 0.0037529408000409603, + 0.005587647669017315, + 0.004845902323722839, + 0.004681260325014591, + 0.005495531018823385, + 0.004307054448872805, + 0.004136842209845781, + 0.005400679539889097, + 0.007498010061681271, + 0.008371527306735516, + 0.008156090043485165, + 0.006302474066615105, + 0.008638673461973667, + 0.007268216926604509, + 0.007183228153735399, + 0.010660052299499512, + 0.012896966189146042, + 0.015065382234752178, + 0.015590612776577473, + 0.0173639003187418, + 0.01980566792190075, + 0.03022881969809532, + 0.03026571311056614, + 0.018316641449928284, + 0.01901177503168583, + 0.025572016835212708, + 0.02208901196718216, + 0.00906411837786436, + 0.010816768743097782, + 0.012051982805132866, + 0.009373112581670284, + 0.02102818712592125, + 0.01633157767355442, + 0.034942395985126495, + 0.022004738450050354, + 0.014880646020174026, + 0.019429555162787437, + 0.04355604574084282, + 0.026409171521663666, + 0.0075786421075463295, + 0.006758011411875486, + 0.009981315582990646, + 0.007349650375545025, + 0.008633890189230442, + 0.003584769321605563, + 0.006977445911616087, + 0.01082164142280817, + 0.010601626709103584, + 0.008040551096200943, + 0.04236965626478195, + 0.010970709845423698, + 0.011050707660615444, + 0.015116149559617043, + 0.03538699075579643, + 0.015215856023132801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.048646602779626846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a low angle, pans rapidly to the right, closely following a modern police car in classic black and white livery. Late in the night, amidst a torrential downpour, the car suddenly enters the frame from the left, speeding off to the right across a wide, tarmacked city street at breakneck speed. Its body, thoroughly drenched by the rain, its glossy surface reflecting everything around it. The red and blue police lights atop the car flash in a rapid, rhythmic pattern, casting a dazzling light across the slippery, mirror-like pavement and the facades of the towering buildings lining the street, creating a constantly chasing, shifting pattern of red and blue. The car's bright headlights pierce the rain, casting twin sharp beams of light, while its rapidly spinning wheels send up streams of shimmering water. In the background, the glow of the city's neon signs and streetlights blends into a patchwork of blurred reflections on the waterlogged road, contrasting with the scattered, warm yellow lights filtering through the windows of distant buildings. The main color tone of the entire picture is cold, dominated by the deep night and cold rain. Only the red and blue police lights and neon colors become jumping highlights, creating a strong sense of urgency, tense atmosphere and dynamic urban rainy night aesthetics.", + "seed": 19255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_0/index_5465_seed_19255_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": -0.03496345710009337, + "vlm_logprob": -0.03496345710009337, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03496345710009337, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052974455058574677 + ], + "opposite_logprob_tokens": [ + -3.8029744625091553 + ], + "logprob_reward": -0.052974455058574677, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.052974455058574677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10068333148956299, + 0.07269857823848724, + 0.056352246552705765, + 0.04684644564986229, + 0.10688284784555435, + 0.1279524266719818, + 0.15797510743141174, + 0.16259771585464478, + 0.0559120737016201, + 0.11209921538829803 + ], + "credit_map": [ + 0.013388816267251968, + 0.009378478862345219, + 0.008841332979500294, + 0.023310886695981026, + 0.007814216427505016, + 0.011098037473857403, + 0.014241619035601616, + 0.012609945610165596, + 0.00911795161664486, + 0.006664689630270004, + 0.007188862655311823, + 0.012360080145299435, + 0.009836191311478615, + 0.011150617152452469, + 0.008109847083687782, + 0.008270341902971268, + 0.00416701752692461, + 0.004783798940479755, + 0.0077667091973125935, + 0.005595823749899864, + 0.005959867499768734, + 0.01318987924605608, + 0.0070256395265460014, + 0.007863515987992287, + 0.0035394513979554176, + 0.005680738016963005, + 0.007116863504052162, + 0.0074570877477526665, + 0.004334151744842529, + 0.005471194162964821, + 0.008373443968594074, + 0.004873516969382763, + 0.011057574301958084, + 0.012709103524684906, + 0.015022488310933113, + 0.017471197992563248, + 0.011216502636671066, + 0.012188960798084736, + 0.013690977357327938, + 0.013526040129363537, + 0.014980265870690346, + 0.014526083134114742, + 0.02274477854371071, + 0.024126142263412476, + 0.011642701923847198, + 0.01431628130376339, + 0.014977478422224522, + 0.010638692416250706, + 0.016337493434548378, + 0.023157523944973946, + 0.022031890228390694, + 0.0347721129655838, + 0.011516076512634754, + 0.026625726372003555, + 0.013212391175329685, + 0.010321902111172676, + 0.02405877597630024, + 0.020319262519478798, + 0.025990260764956474, + 0.023653095588088036, + 0.02089243195950985, + 0.01668296381831169, + 0.020751090720295906, + 0.01024982612580061, + 0.009434509091079235, + 0.005560318473726511, + 0.007783368229866028, + 0.005371470469981432, + 0.0053736865520477295, + 0.007864238694310188, + 0.008182360790669918, + 0.006342122796922922, + 0.010107364505529404, + 0.00773023022338748, + 0.00731938611716032, + 0.01002469938248396, + 0.00650656595826149, + 0.011949882842600346, + 0.03208994120359421, + 0.026371151208877563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.052974455058574677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374120593070984 + ], + "opposite_logprob_tokens": [ + -3.4087412357330322 + ], + "logprob_reward": -0.03374120593070984, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374120593070984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07690689712762833, + 0.10333412140607834, + 0.07770685106515884, + 0.02054072730243206, + 0.07369839400053024, + 0.056179653853178024, + 0.1906823366880417, + 0.11907827109098434, + 0.20696420967578888, + 0.0749085322022438 + ], + "credit_map": [ + 0.013731135986745358, + 0.012982608750462532, + 0.007551861926913261, + 0.012215853668749332, + 0.005305795930325985, + 0.006882543209940195, + 0.007925980724394321, + 0.01031111553311348, + 0.011008222587406635, + 0.008266965858638287, + 0.0075750467367470264, + 0.012309905141592026, + 0.01588396169245243, + 0.019504845142364502, + 0.016806233674287796, + 0.01197894848883152, + 0.010163117200136185, + 0.0145694799721241, + 0.007220132742077112, + 0.008624810725450516, + 0.009423990733921528, + 0.00785620603710413, + 0.00982338935136795, + 0.010025719180703163, + 0.001356373424641788, + 0.002210627542808652, + 0.0032495649065822363, + 0.0020771301351487637, + 0.002751627005636692, + 0.002872211392968893, + 0.0030245468951761723, + 0.0029986444860696793, + 0.008132320828735828, + 0.010789277032017708, + 0.01020501833409071, + 0.013659954071044922, + 0.00671609491109848, + 0.009650428779423237, + 0.007353495806455612, + 0.007191803772002459, + 0.004934654571115971, + 0.006364442408084869, + 0.009528450667858124, + 0.011956397444009781, + 0.007464712020009756, + 0.0060571106150746346, + 0.005910275038331747, + 0.0039636120200157166, + 0.026865599676966667, + 0.020243685692548752, + 0.02684657834470272, + 0.04086478799581528, + 0.01428027544170618, + 0.037824925035238266, + 0.013177203945815563, + 0.010579287074506283, + 0.013433928601443768, + 0.0160001739859581, + 0.016441108658909798, + 0.017901189625263214, + 0.01203804649412632, + 0.014671465381979942, + 0.01603047363460064, + 0.012561891227960587, + 0.027275165542960167, + 0.031042149290442467, + 0.023280153051018715, + 0.013995109125971794, + 0.03210923820734024, + 0.037464264780282974, + 0.026454932987689972, + 0.015343195758759975, + 0.008325217291712761, + 0.006212235428392887, + 0.008342098444700241, + 0.006335054058581591, + 0.004698139615356922, + 0.008519028313457966, + 0.018708789721131325, + 0.013767972588539124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.03374120593070984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020605430006980896 + ], + "opposite_logprob_tokens": [ + -3.8956053256988525 + ], + "logprob_reward": -0.020605430006980896, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020605430006980896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1529047191143036, + 0.08602037280797958, + 0.05067111551761627, + 0.029971757903695107, + 0.07203035801649094, + 0.05147630348801613, + 0.162303164601326, + 0.04286928474903107, + 0.21911489963531494, + 0.13263806700706482 + ], + "credit_map": [ + 0.01653822511434555, + 0.013930109329521656, + 0.02811369299888611, + 0.03396311774849892, + 0.01096282433718443, + 0.012797283008694649, + 0.0174313485622406, + 0.01916811615228653, + 0.00706033268943429, + 0.007867075502872467, + 0.006647549103945494, + 0.012218451127409935, + 0.010935020633041859, + 0.01833416521549225, + 0.011812726967036724, + 0.011145039461553097, + 0.004865836817771196, + 0.004436397459357977, + 0.006384833715856075, + 0.00613720528781414, + 0.007592386566102505, + 0.007166014518588781, + 0.0075982194393873215, + 0.006490218453109264, + 0.0026614046655595303, + 0.003585780505090952, + 0.004357283003628254, + 0.00350987003184855, + 0.004147135652601719, + 0.0037253189366310835, + 0.003936734050512314, + 0.004048227798193693, + 0.006172258406877518, + 0.009299901314079762, + 0.010703089646995068, + 0.01349592674523592, + 0.00771211227402091, + 0.010668016038835049, + 0.008012677542865276, + 0.0059663718566298485, + 0.00555800274014473, + 0.007027982734143734, + 0.007579387631267309, + 0.008097087033092976, + 0.007759518455713987, + 0.004952207673341036, + 0.005728230811655521, + 0.004773883149027824, + 0.0191744863986969, + 0.022256648167967796, + 0.019055945798754692, + 0.03676798567175865, + 0.0158392246812582, + 0.0256804171949625, + 0.015396724455058575, + 0.008131715469062328, + 0.004244402050971985, + 0.006604246329516172, + 0.00639326823875308, + 0.0066668190993368626, + 0.006179718766361475, + 0.005073126871138811, + 0.004839354194700718, + 0.0028683484997600317, + 0.028610961511731148, + 0.027306873351335526, + 0.024429183453321457, + 0.01686646230518818, + 0.031109021976590157, + 0.04538454860448837, + 0.027388911694288254, + 0.018018921837210655, + 0.015234011225402355, + 0.008976008743047714, + 0.012064186856150627, + 0.008932370692491531, + 0.008031207136809826, + 0.012936735525727272, + 0.03725530579686165, + 0.029208239167928696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.020605430006980896 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15224631130695343, + 0.06808331608772278, + 0.06355597078800201, + 0.024929139763116837, + 0.03603168949484825, + 0.02485257014632225, + 0.038155630230903625, + 0.061950910836458206, + 0.27359840273857117, + 0.25659602880477905 + ], + "credit_map": [ + 0.020789818838238716, + 0.02390504628419876, + 0.010442079044878483, + 0.01894991099834442, + 0.016031095758080482, + 0.021069249138236046, + 0.018070582300424576, + 0.022988516837358475, + 0.0067659104242920876, + 0.008010737597942352, + 0.004913329612463713, + 0.009068377315998077, + 0.010335155762732029, + 0.010182347148656845, + 0.010303797200322151, + 0.008503654971718788, + 0.007741580251604319, + 0.010396236553788185, + 0.0046871621161699295, + 0.006728175096213818, + 0.00753912003710866, + 0.009304867126047611, + 0.007556966505944729, + 0.009601865895092487, + 0.002482140436768532, + 0.002892811084166169, + 0.004839145112782717, + 0.0026171330828219652, + 0.0024978462606668472, + 0.0031031363178044558, + 0.002090063877403736, + 0.004406863823533058, + 0.004124983213841915, + 0.0035105308052152395, + 0.004212723113596439, + 0.0036174440756440163, + 0.005468444898724556, + 0.008793079294264317, + 0.003049355000257492, + 0.0032551304902881384, + 0.002202394185587764, + 0.002794391242787242, + 0.003947680350393057, + 0.004057608544826508, + 0.0037840050645172596, + 0.0023635528050363064, + 0.0029709767550230026, + 0.0027319611981511116, + 0.004266176838427782, + 0.006485311780124903, + 0.005621833261102438, + 0.006500854156911373, + 0.003966043703258038, + 0.006358680780977011, + 0.0028519255574792624, + 0.0021048029884696007, + 0.007348346523940563, + 0.010478978976607323, + 0.008118686266243458, + 0.007805619854480028, + 0.010637414641678333, + 0.006631521508097649, + 0.007898900657892227, + 0.0030314382165670395, + 0.042726971209049225, + 0.0425109937787056, + 0.0355319045484066, + 0.018290039151906967, + 0.040985383093357086, + 0.035983577370643616, + 0.03695879876613617, + 0.0206107459962368, + 0.0335286483168602, + 0.021141292527318, + 0.034845851361751556, + 0.01917167566716671, + 0.01699860207736492, + 0.031066028401255608, + 0.056256651878356934, + 0.043587274849414825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549527183175087 + ], + "opposite_logprob_tokens": [ + -2.929952621459961 + ], + "logprob_reward": -0.0549527183175087, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549527183175087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037991877645254135, + 0.16733478009700775, + 0.10439900308847427, + 0.02138776145875454, + 0.03843642771244049, + 0.05686740204691887, + 0.28717896342277527, + 0.03489544987678528, + 0.21378730237483978, + 0.03772103786468506 + ], + "credit_map": [ + 0.00404709717258811, + 0.0037631248123943806, + 0.0038512269966304302, + 0.005161553621292114, + 0.004641341045498848, + 0.004227812401950359, + 0.006169881206005812, + 0.006129843648523092, + 0.012675936333835125, + 0.012214378453791142, + 0.012378595769405365, + 0.015309476293623447, + 0.01838826574385166, + 0.04837764427065849, + 0.024067632853984833, + 0.023922869935631752, + 0.007752608973532915, + 0.008399313315749168, + 0.01172943040728569, + 0.011788913048803806, + 0.01404243241995573, + 0.012615957297384739, + 0.021754875779151917, + 0.016315478831529617, + 0.001880436553619802, + 0.0027006035670638084, + 0.0030119908042252064, + 0.0017246975330635905, + 0.0024777527432888746, + 0.003642195137217641, + 0.0019571399316191673, + 0.003992946818470955, + 0.004198768641799688, + 0.004794853273779154, + 0.005300972610712051, + 0.005425490904599428, + 0.004824216943234205, + 0.006198658142238855, + 0.003559173783287406, + 0.004134297836571932, + 0.004972878843545914, + 0.00708083575591445, + 0.008988172747194767, + 0.01084277592599392, + 0.008190460503101349, + 0.005273790564388037, + 0.0063163600862026215, + 0.005202132277190685, + 0.0410427488386631, + 0.04205642640590668, + 0.036938123404979706, + 0.06856780499219894, + 0.023195641115307808, + 0.04072852432727814, + 0.02160659246146679, + 0.013043113052845001, + 0.004968092776834965, + 0.002895460231229663, + 0.004899299703538418, + 0.005414357874542475, + 0.003336701774969697, + 0.005546084139496088, + 0.004400352016091347, + 0.0034351013600826263, + 0.028737496584653854, + 0.03231179714202881, + 0.024293078109622, + 0.015004941262304783, + 0.03269437700510025, + 0.036628372967243195, + 0.02838936075568199, + 0.01572788693010807, + 0.004085474647581577, + 0.004423687234520912, + 0.004788077436387539, + 0.004746868275105953, + 0.0031141452491283417, + 0.004458188079297543, + 0.007021503057330847, + 0.005083097610622644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.0549527183175087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a low angle, pans rapidly to the right, closely following a modern police car in classic black and white livery. Late in the night, amidst a torrential downpour, the car suddenly enters the frame from the left, speeding off to the right across a wide, tarmacked city street at breakneck speed. Its body, thoroughly drenched by the rain, its glossy surface reflecting everything around it. The red and blue police lights atop the car flash in a rapid, rhythmic pattern, casting a dazzling light across the slippery, mirror-like pavement and the facades of the towering buildings lining the street, creating a constantly chasing, shifting pattern of red and blue. The car's bright headlights pierce the rain, casting twin sharp beams of light, while its rapidly spinning wheels send up streams of shimmering water. In the background, the glow of the city's neon signs and streetlights blends into a patchwork of blurred reflections on the waterlogged road, contrasting with the scattered, warm yellow lights filtering through the windows of distant buildings. The main color tone of the entire picture is cold, dominated by the deep night and cold rain. Only the red and blue police lights and neon colors become jumping highlights, creating a strong sense of urgency, tense atmosphere and dynamic urban rainy night aesthetics.", + "seed": 19256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_0/index_5465_seed_19256_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": -0.05584526360034943, + "vlm_logprob": -0.05584526360034943, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.05584526360034943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09296347200870514 + ], + "opposite_logprob_tokens": [ + -2.592963457107544 + ], + "logprob_reward": -0.09296347200870514, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09296347200870514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06415097415447235, + 0.03917374834418297, + 0.028998592868447304, + 0.02963951602578163, + 0.06994903087615967, + 0.19014997780323029, + 0.3382539451122284, + 0.037467800080776215, + 0.019264966249465942, + 0.1829514354467392 + ], + "credit_map": [ + 0.00853047240525484, + 0.009160925634205341, + 0.007153516169637442, + 0.005466499365866184, + 0.007538952864706516, + 0.009614713490009308, + 0.00639966269955039, + 0.010286231525242329, + 0.00277679692953825, + 0.004019428510218859, + 0.0021056863479316235, + 0.0026172809302806854, + 0.007449904456734657, + 0.008057455532252789, + 0.006496260408312082, + 0.005650937091559172, + 0.002909150905907154, + 0.0025683792773634195, + 0.002326002810150385, + 0.0030142657924443483, + 0.005942229181528091, + 0.004555033054202795, + 0.0045292978174984455, + 0.003154236590489745, + 0.0023411945439875126, + 0.0033145358320325613, + 0.003468225011602044, + 0.0045480914413928986, + 0.0036143632605671883, + 0.003192868083715439, + 0.005461510270833969, + 0.003698725951835513, + 0.004305428825318813, + 0.002949889749288559, + 0.005284814164042473, + 0.004111285787075758, + 0.0040006255730986595, + 0.022804131731390953, + 0.004584122449159622, + 0.021908733993768692, + 0.008125151507556438, + 0.007679823320358992, + 0.06612211465835571, + 0.028124788776040077, + 0.01816696859896183, + 0.011508002877235413, + 0.009165684692561626, + 0.04125743731856346, + 0.008185356855392456, + 0.011565028689801693, + 0.013480212539434433, + 0.01876932568848133, + 0.011968261562287807, + 0.12981556355953217, + 0.01658603549003601, + 0.127884179353714, + 0.0024341729003936052, + 0.004793226253241301, + 0.005827492568641901, + 0.004448711406439543, + 0.00360090471804142, + 0.0045110792852938175, + 0.006900138687342405, + 0.004952074959874153, + 0.0014911176403984427, + 0.0021120631136000156, + 0.003592785680666566, + 0.0038006315007805824, + 0.001965974923223257, + 0.0012989117531105876, + 0.0028302334249019623, + 0.002173246815800667, + 0.013483872637152672, + 0.03757987543940544, + 0.024540847167372704, + 0.027111053466796875, + 0.03199470043182373, + 0.01461823284626007, + 0.020644646137952805, + 0.012978212907910347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.09296347200870514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790383517742157 + ], + "opposite_logprob_tokens": [ + -2.579038381576538 + ], + "logprob_reward": -0.0790383517742157, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790383517742157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08492137491703033, + 0.11772817373275757, + 0.170253723859787, + 0.11431081593036652, + 0.07337481528520584, + 0.07528208941221237, + 0.17870748043060303, + 0.03850005567073822, + 0.024049172177910805, + 0.12287233769893646 + ], + "credit_map": [ + 0.008913052268326283, + 0.012960786931216717, + 0.00988055020570755, + 0.006593411788344383, + 0.007766667753458023, + 0.014188796281814575, + 0.006948503665626049, + 0.01766960509121418, + 0.00899429526180029, + 0.009405569173395634, + 0.009808198548853397, + 0.0073322756215929985, + 0.026923537254333496, + 0.01836247183382511, + 0.01832578517496586, + 0.018576042726635933, + 0.014640843495726585, + 0.01670764945447445, + 0.009626226499676704, + 0.011184915900230408, + 0.034705404192209244, + 0.03499621897935867, + 0.028682604432106018, + 0.0197098720818758, + 0.009458497166633606, + 0.011851143091917038, + 0.01409170962870121, + 0.01668451353907585, + 0.016074465587735176, + 0.014116315171122551, + 0.017338987439870834, + 0.014695186167955399, + 0.006948533933609724, + 0.004828006960451603, + 0.012642745859920979, + 0.008881846442818642, + 0.00900219939649105, + 0.011839723214507103, + 0.011600774712860584, + 0.007630990352481604, + 0.0035392274148762226, + 0.0064239404164254665, + 0.019133100286126137, + 0.010983394458889961, + 0.009780988097190857, + 0.007165446877479553, + 0.006697750650346279, + 0.011558243073523045, + 0.005156850442290306, + 0.009756597690284252, + 0.010689529590308666, + 0.011569247581064701, + 0.009999583475291729, + 0.06111959367990494, + 0.010981056839227676, + 0.059435002505779266, + 0.0022750343196094036, + 0.0031708639580756426, + 0.004668089561164379, + 0.005093884188681841, + 0.005102639552205801, + 0.0058901188895106316, + 0.008019420318305492, + 0.004280003719031811, + 0.001661064918152988, + 0.003455575555562973, + 0.003769531147554517, + 0.0042799729853868484, + 0.00253677717410028, + 0.002392763737589121, + 0.0028315105009824038, + 0.0031219765078276396, + 0.009996121749281883, + 0.025137756019830704, + 0.018112143501639366, + 0.017913592979311943, + 0.019500020891427994, + 0.00936334952712059, + 0.013379826210439205, + 0.009469516575336456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.0790383517742157 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609399303793907 + ], + "opposite_logprob_tokens": [ + -3.8956093788146973 + ], + "logprob_reward": -0.020609399303793907, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609399303793907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062425192445516586, + 0.05688634514808655, + 0.05552482604980469, + 0.17026032507419586, + 0.15140685439109802, + 0.28289806842803955, + 0.0345926433801651, + 0.0320948101580143, + 0.03497365862131119, + 0.11893728375434875 + ], + "credit_map": [ + 0.006155727431178093, + 0.0072351559065282345, + 0.006832306273281574, + 0.006702072452753782, + 0.008102656342089176, + 0.008382431231439114, + 0.007401719223707914, + 0.011613121256232262, + 0.0061258370988070965, + 0.007668077480047941, + 0.004732102621346712, + 0.00429524015635252, + 0.006907735485583544, + 0.008837025612592697, + 0.01081058755517006, + 0.007509738206863403, + 0.005449949763715267, + 0.005826571490615606, + 0.003239416517317295, + 0.005520407576113939, + 0.011632949113845825, + 0.01117852795869112, + 0.007384362630546093, + 0.005292639601975679, + 0.014457053504884243, + 0.01325839664787054, + 0.02173476293683052, + 0.03188567981123924, + 0.017609853297472, + 0.017052583396434784, + 0.03334890305995941, + 0.02091309428215027, + 0.007561868522316217, + 0.004824937786906958, + 0.01771288365125656, + 0.00716385105624795, + 0.009547392837703228, + 0.0552600659430027, + 0.009814152494072914, + 0.039521701633930206, + 0.011246477253735065, + 0.012919289991259575, + 0.09950674325227737, + 0.0473727211356163, + 0.0249783918261528, + 0.022603347897529602, + 0.011205642484128475, + 0.0530654639005661, + 0.003742365399375558, + 0.005110753700137138, + 0.005947272293269634, + 0.005391263402998447, + 0.003751541255041957, + 0.0036840522661805153, + 0.004376217257231474, + 0.0025891773402690887, + 0.002097255550324917, + 0.003696393221616745, + 0.00346062402240932, + 0.004974632989615202, + 0.0045005823485553265, + 0.004339275881648064, + 0.005154857411980629, + 0.0038711903616786003, + 0.0028064216021448374, + 0.003435431746765971, + 0.006320958957076073, + 0.00475835707038641, + 0.003221650142222643, + 0.0032705955673009157, + 0.008425532840192318, + 0.0027347116265445948, + 0.015558000653982162, + 0.015912489965558052, + 0.01881093531847, + 0.018718840554356575, + 0.012859033420681953, + 0.014892008155584335, + 0.01181864645332098, + 0.010367332957684994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.020609399303793907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.412618488073349, + 0.03202086687088013, + 0.028682874515652657, + 0.0367535762488842, + 0.02219162881374359, + 0.16975513100624084, + 0.07549048960208893, + 0.05013260990381241, + 0.03748311102390289, + 0.13487115502357483 + ], + "credit_map": [ + 0.03401486203074455, + 0.05253545194864273, + 0.04036397114396095, + 0.02905561774969101, + 0.04305107146501541, + 0.08314389735460281, + 0.038945138454437256, + 0.09150852262973785, + 0.0035012210719287395, + 0.003020194126293063, + 0.00321925338357687, + 0.0030638426542282104, + 0.004157539922744036, + 0.004940972663462162, + 0.005887285340577364, + 0.004230561200529337, + 0.001979381311684847, + 0.0021882487926632166, + 0.0012853399384766817, + 0.001703667570836842, + 0.009293868206441402, + 0.004320324398577213, + 0.005262656603008509, + 0.0026493892073631287, + 0.004535730928182602, + 0.004422896541655064, + 0.0032199460547417402, + 0.003162963315844536, + 0.003962352406233549, + 0.0031824877951294184, + 0.00904294103384018, + 0.005224260035902262, + 0.001845778664574027, + 0.0016865398501977324, + 0.004133128095418215, + 0.002183429431170225, + 0.002698473632335663, + 0.0032273458782583475, + 0.0033262656070291996, + 0.0030906691681593657, + 0.007346784695982933, + 0.007754625286906958, + 0.06759359687566757, + 0.03174318000674248, + 0.006352019961923361, + 0.006456415168941021, + 0.004790246486663818, + 0.037718288600444794, + 0.0035934876650571823, + 0.0041402000933885574, + 0.006411922164261341, + 0.005181684624403715, + 0.0038370406255126, + 0.022845683619379997, + 0.004843111615628004, + 0.024637367576360703, + 0.0029854029417037964, + 0.005659348797053099, + 0.006500823423266411, + 0.005936052184551954, + 0.006854488048702478, + 0.00788091029971838, + 0.008943988010287285, + 0.0053716013208031654, + 0.0029281710740178823, + 0.002769716549664736, + 0.00659424951300025, + 0.0056825377978384495, + 0.003641018643975258, + 0.003814836498349905, + 0.007508486043661833, + 0.004544100258499384, + 0.011667376384139061, + 0.029469985514879227, + 0.018344467505812645, + 0.0228880662471056, + 0.019763793796300888, + 0.009371497668325901, + 0.013183210976421833, + 0.010182764381170273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.007633799687027931 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898129522800446 + ], + "opposite_logprob_tokens": [ + -2.578981399536133 + ], + "logprob_reward": -0.07898129522800446, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898129522800446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07149308919906616, + 0.043784819543361664, + 0.04647484049201012, + 0.1641678661108017, + 0.17359288036823273, + 0.05411127954721451, + 0.21974551677703857, + 0.07592159509658813, + 0.023243175819516182, + 0.12746499478816986 + ], + "credit_map": [ + 0.006937825120985508, + 0.012319677509367466, + 0.006783881690353155, + 0.005257174372673035, + 0.010094731114804745, + 0.013238116167485714, + 0.006186122540384531, + 0.010675548575818539, + 0.004326701629906893, + 0.003786669112741947, + 0.0038093016482889652, + 0.004180246032774448, + 0.009116766043007374, + 0.006185711361467838, + 0.005823933053761721, + 0.006555488798767328, + 0.005361249204725027, + 0.004310331307351589, + 0.004428125452250242, + 0.005170042626559734, + 0.007864833809435368, + 0.00622550630941987, + 0.006231855601072311, + 0.0068828933872282505, + 0.018399441614747047, + 0.019443189725279808, + 0.012495085597038269, + 0.01570042222738266, + 0.018261780962347984, + 0.012158490717411041, + 0.047147348523139954, + 0.020562095567584038, + 0.0075587076134979725, + 0.008976763114333153, + 0.024626825004816055, + 0.008174465037882328, + 0.017239706590771675, + 0.05920111760497093, + 0.011595949530601501, + 0.03621932864189148, + 0.00206207693554461, + 0.0026620132848620415, + 0.013977625407278538, + 0.005903381854295731, + 0.004373615141957998, + 0.009445183910429478, + 0.003237438155338168, + 0.012449943460524082, + 0.010135277174413204, + 0.014929673634469509, + 0.016247177496552467, + 0.01925361342728138, + 0.01866380125284195, + 0.06188581883907318, + 0.02256850339472294, + 0.05606162175536156, + 0.004381146281957626, + 0.005093595013022423, + 0.007370135746896267, + 0.009867041371762753, + 0.009724519215524197, + 0.01260718796402216, + 0.01980622112751007, + 0.007071744184941053, + 0.0020891083404421806, + 0.0021776510402560234, + 0.004286202136427164, + 0.004060311242938042, + 0.0018297246424481273, + 0.0024142430629581213, + 0.004571450874209404, + 0.0018144809873774648, + 0.013215987011790276, + 0.017109200358390808, + 0.018263092264533043, + 0.020297931507229805, + 0.016447799280285835, + 0.019514648243784904, + 0.01357234362512827, + 0.009043981321156025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.07898129522800446 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic tracking shot, shot from a low angle, pans rapidly to the right, closely following a modern police car in classic black and white livery. Late in the night, amidst a torrential downpour, the car suddenly enters the frame from the left, speeding off to the right across a wide, tarmacked city street at breakneck speed. Its body, thoroughly drenched by the rain, its glossy surface reflecting everything around it. The red and blue police lights atop the car flash in a rapid, rhythmic pattern, casting a dazzling light across the slippery, mirror-like pavement and the facades of the towering buildings lining the street, creating a constantly chasing, shifting pattern of red and blue. The car's bright headlights pierce the rain, casting twin sharp beams of light, while its rapidly spinning wheels send up streams of shimmering water. In the background, the glow of the city's neon signs and streetlights blends into a patchwork of blurred reflections on the waterlogged road, contrasting with the scattered, warm yellow lights filtering through the windows of distant buildings. The main color tone of the entire picture is cold, dominated by the deep night and cold rain. Only the red and blue police lights and neon colors become jumping highlights, creating a strong sense of urgency, tense atmosphere and dynamic urban rainy night aesthetics.", + "seed": 19257, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_0/index_5465_seed_19257_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": -0.07102025672793388, + "vlm_logprob": -0.07102025672793388, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.07102025672793388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09911051392555237 + ], + "opposite_logprob_tokens": [ + -2.5991106033325195 + ], + "logprob_reward": -0.09911051392555237, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09911051392555237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06687968969345093, + 0.0809907615184784, + 0.06537919491529465, + 0.06808945536613464, + 0.11798551678657532, + 0.1353815197944641, + 0.23372553288936615, + 0.09206521511077881, + 0.08999839425086975, + 0.04950473830103874 + ], + "credit_map": [ + 0.009443272836506367, + 0.010142043232917786, + 0.006374835502356291, + 0.00490185059607029, + 0.008189409039914608, + 0.010057882405817509, + 0.007664923556149006, + 0.010105477645993233, + 0.005166943650692701, + 0.010183245874941349, + 0.007380176801234484, + 0.009097068570554256, + 0.009001968428492546, + 0.012228411622345448, + 0.014270481653511524, + 0.013662466779351234, + 0.005748607683926821, + 0.011560877785086632, + 0.008752488531172276, + 0.008784000761806965, + 0.007787523325532675, + 0.0076679568737745285, + 0.008479299023747444, + 0.006598435342311859, + 0.007765474263578653, + 0.010435295291244984, + 0.0052043781615793705, + 0.0046700770035386086, + 0.012026671320199966, + 0.004721074365079403, + 0.008906482718884945, + 0.014359998516738415, + 0.020653357729315758, + 0.010327504016458988, + 0.015968138352036476, + 0.031412579119205475, + 0.009575656615197659, + 0.006366343703120947, + 0.012716841883957386, + 0.010965093038976192, + 0.01586236245930195, + 0.013527589850127697, + 0.027666529640555382, + 0.019467204809188843, + 0.014564655721187592, + 0.010589764453470707, + 0.013087145052850246, + 0.02061626873910427, + 0.010659909807145596, + 0.013508280739188194, + 0.013055684976279736, + 0.06271575391292572, + 0.019977018237113953, + 0.019414817914366722, + 0.03621277958154678, + 0.058181293308734894, + 0.005359519738703966, + 0.008661042898893356, + 0.005630956031382084, + 0.01847449317574501, + 0.011204775422811508, + 0.009725119918584824, + 0.017162395641207695, + 0.01584690436720848, + 0.01078813523054123, + 0.014594510197639465, + 0.01376790925860405, + 0.013323660008609295, + 0.010279858484864235, + 0.009625089354813099, + 0.010621210560202599, + 0.006998022552579641, + 0.006100064609199762, + 0.005276921670883894, + 0.00785837508738041, + 0.006643995642662048, + 0.0064606377854943275, + 0.0059303841553628445, + 0.005397969391196966, + 0.005836389493197203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.09911051392555237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129976138472557 + ], + "opposite_logprob_tokens": [ + -2.237997531890869 + ], + "logprob_reward": -0.1129976138472557, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129976138472557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17689059674739838, + 0.14652180671691895, + 0.08231174200773239, + 0.05249812453985214, + 0.10080335289239883, + 0.09727195650339127, + 0.036263711750507355, + 0.07778529077768326, + 0.17304560542106628, + 0.05660783126950264 + ], + "credit_map": [ + 0.015581646002829075, + 0.014352457597851753, + 0.013957002200186253, + 0.008408322930335999, + 0.02273312583565712, + 0.0372670479118824, + 0.01671651564538479, + 0.04787446931004524, + 0.012348830699920654, + 0.01036021113395691, + 0.007868689484894276, + 0.005388977471739054, + 0.008371295407414436, + 0.04558642581105232, + 0.007598471362143755, + 0.04899892583489418, + 0.006502375937998295, + 0.01075089443475008, + 0.013156800530850887, + 0.01191692054271698, + 0.008196023292839527, + 0.011465593241155148, + 0.009352447465062141, + 0.010970689356327057, + 0.004837690386921167, + 0.005839543417096138, + 0.0030605450738221407, + 0.004119719844311476, + 0.011893017217516899, + 0.005130849778652191, + 0.006715389434248209, + 0.010901368223130703, + 0.01740659959614277, + 0.006517802365124226, + 0.01118344534188509, + 0.0353035144507885, + 0.006235102657228708, + 0.005112277343869209, + 0.011069713160395622, + 0.007974897511303425, + 0.008053301833570004, + 0.009124226868152618, + 0.014894293621182442, + 0.02196260169148445, + 0.01061534509062767, + 0.005881907884031534, + 0.011377709917724133, + 0.015362569130957127, + 0.0031426132190972567, + 0.0040053995326161385, + 0.003341471776366234, + 0.005624887067824602, + 0.003927294630557299, + 0.004917359910905361, + 0.0051748985424637794, + 0.006129786837846041, + 0.004784667864441872, + 0.008349105715751648, + 0.004450235981494188, + 0.014269214123487473, + 0.006632596720010042, + 0.011848416179418564, + 0.012396074831485748, + 0.015054982155561447, + 0.037609245628118515, + 0.025484643876552582, + 0.02930356375873089, + 0.023310914635658264, + 0.01349527295678854, + 0.01499080378562212, + 0.013925766572356224, + 0.014925392344594002, + 0.005478886887431145, + 0.005046119447797537, + 0.010067196562886238, + 0.01045612059533596, + 0.0038476232439279556, + 0.005943450145423412, + 0.0062256152741611, + 0.009542817249894142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.1129976138472557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023321032524108887 + ], + "opposite_logprob_tokens": [ + -3.7733211517333984 + ], + "logprob_reward": -0.023321032524108887, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023321032524108887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1787199229001999, + 0.1744353324174881, + 0.05941136181354523, + 0.07104773819446564, + 0.05369722843170166, + 0.10911707580089569, + 0.10653185844421387, + 0.06996765732765198, + 0.06463304162025452, + 0.11243873834609985 + ], + "credit_map": [ + 0.026599617674946785, + 0.01706388033926487, + 0.011495268903672695, + 0.01013509277254343, + 0.030855340883135796, + 0.03075706958770752, + 0.014006702229380608, + 0.03780695050954819, + 0.01039687730371952, + 0.012352482415735722, + 0.009982218034565449, + 0.011560718528926373, + 0.014047721400856972, + 0.04970739781856537, + 0.015602792613208294, + 0.050785113126039505, + 0.008793861605226994, + 0.006204951088875532, + 0.008893907070159912, + 0.006467796862125397, + 0.005962397903203964, + 0.007639092858880758, + 0.007292938884347677, + 0.008156418800354004, + 0.0067387353628873825, + 0.010784110985696316, + 0.00393688166514039, + 0.00405854033306241, + 0.013548911549150944, + 0.0072515797801315784, + 0.007095578592270613, + 0.01763339899480343, + 0.008957651443779469, + 0.003901305142790079, + 0.014846217818558216, + 0.00552730169147253, + 0.003998355008661747, + 0.004622865468263626, + 0.005452336277812719, + 0.006391195114701986, + 0.009472611360251904, + 0.00979952234774828, + 0.020294642075896263, + 0.024639563634991646, + 0.008922329172492027, + 0.005429356824606657, + 0.013577418401837349, + 0.016981640830636024, + 0.004604618530720472, + 0.006490684114396572, + 0.0051863226108253, + 0.03361362963914871, + 0.006028459873050451, + 0.007438070606440306, + 0.00825461931526661, + 0.03491545468568802, + 0.007605098187923431, + 0.012520289048552513, + 0.004021863453090191, + 0.010084873996675014, + 0.00982936006039381, + 0.006818041671067476, + 0.010899157263338566, + 0.008188975974917412, + 0.010596413165330887, + 0.007448532152920961, + 0.010657875798642635, + 0.007583618629723787, + 0.006786273326724768, + 0.006180086638778448, + 0.007953428663313389, + 0.007426809519529343, + 0.012221262790262699, + 0.011763458140194416, + 0.01943584531545639, + 0.017364511266350746, + 0.012074376456439495, + 0.008719397708773613, + 0.01825082302093506, + 0.012609072029590607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.023321032524108887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067422278225421906 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.0067422278225421906, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067422278225421906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2756659686565399, + 0.10863254219293594, + 0.0571431890130043, + 0.05441785231232643, + 0.04297366738319397, + 0.07540405541658401, + 0.06568416208028793, + 0.0556940883398056, + 0.2215452641248703, + 0.04283924773335457 + ], + "credit_map": [ + 0.0196889229118824, + 0.015265862457454205, + 0.019723201170563698, + 0.011584065854549408, + 0.03687266632914543, + 0.060545407235622406, + 0.022092197090387344, + 0.08989366888999939, + 0.01029053796082735, + 0.013323551043868065, + 0.009703789837658405, + 0.011196295730769634, + 0.0115501768887043, + 0.02072567492723465, + 0.013935263268649578, + 0.01790725067257881, + 0.004448976367712021, + 0.00596308708190918, + 0.008805968798696995, + 0.010273516178131104, + 0.006366146262735128, + 0.00593927875161171, + 0.008619535714387894, + 0.006726682186126709, + 0.003531989874318242, + 0.005693850107491016, + 0.003194260410964489, + 0.004106428939849138, + 0.01358521543443203, + 0.006415796000510454, + 0.004820146132260561, + 0.013070164248347282, + 0.006019903346896172, + 0.003267405554652214, + 0.0054869758896529675, + 0.011779790744185448, + 0.00271984632126987, + 0.00353762018494308, + 0.005381837021559477, + 0.004780289251357317, + 0.01418053638190031, + 0.009560992941260338, + 0.012035899795591831, + 0.010673516429960728, + 0.00500987796112895, + 0.005735797807574272, + 0.006647788919508457, + 0.011559649370610714, + 0.0036113227251917124, + 0.0043580131605267525, + 0.0042661200277507305, + 0.019260257482528687, + 0.003103273455053568, + 0.0038345232605934143, + 0.005929482169449329, + 0.021321162581443787, + 0.003481696592643857, + 0.010240614414215088, + 0.0037570479325950146, + 0.012736967764794827, + 0.0033467928878962994, + 0.005268540233373642, + 0.006601417902857065, + 0.010261011309921741, + 0.05458638817071915, + 0.027919305488467216, + 0.037761103361845016, + 0.02822255529463291, + 0.0198472049087286, + 0.016141638159751892, + 0.017578408122062683, + 0.019488655030727386, + 0.004348489455878735, + 0.0057571264915168285, + 0.008018625900149345, + 0.006951602641493082, + 0.0031925903167575598, + 0.003947329707443714, + 0.0041880919598042965, + 0.006435389164835215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0067422278225421906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern police car with classic black and white livery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car enter from the left and speed off to the right across a wide tarmacked city street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are red and blue police lights flashing on the car in a rapid, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set late at night during a torrential downpour with a low-angle camera shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan rapidly to the right while closely following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.34746432304382324, + 0.06293909251689911, + 0.051279518753290176, + 0.09818731248378754, + 0.04042559117078781, + 0.055567894130945206, + 0.13361313939094543, + 0.06711815297603607, + 0.10643579065799713, + 0.03696919232606888 + ], + "credit_map": [ + 0.046379510313272476, + 0.031967464834451675, + 0.028592150658369064, + 0.014008037745952606, + 0.04441928118467331, + 0.0653073638677597, + 0.028688840568065643, + 0.08810165524482727, + 0.006329767405986786, + 0.005776617210358381, + 0.0041468506678938866, + 0.004208657890558243, + 0.004495791159570217, + 0.014123045839369297, + 0.008291684091091156, + 0.01556667871773243, + 0.00425930880010128, + 0.005020348355174065, + 0.005823836196213961, + 0.005769222043454647, + 0.007605572231113911, + 0.006910684052854776, + 0.00818176381289959, + 0.007708781864494085, + 0.007569200359284878, + 0.012155258096754551, + 0.005163934081792831, + 0.005400488153100014, + 0.02432313933968544, + 0.00814752746373415, + 0.00889639277011156, + 0.026531372219324112, + 0.006785160396248102, + 0.004586227238178253, + 0.008051413111388683, + 0.005722166039049625, + 0.0039237067103385925, + 0.0030510909855365753, + 0.003525031730532646, + 0.004780797753483057, + 0.007288412656635046, + 0.008186249993741512, + 0.007962848991155624, + 0.006954704876989126, + 0.004674867261201143, + 0.004557549487799406, + 0.00684761768206954, + 0.009095638990402222, + 0.00675083277747035, + 0.006136439740657806, + 0.004564385395497084, + 0.04397337883710861, + 0.005665032658725977, + 0.008850022219121456, + 0.010209040716290474, + 0.04746400937438011, + 0.003089782316237688, + 0.010872084647417068, + 0.0036429334431886673, + 0.017751997336745262, + 0.004243756178766489, + 0.005153804551810026, + 0.010370777919888496, + 0.011993013322353363, + 0.029188431799411774, + 0.010164213366806507, + 0.017754795029759407, + 0.014179158955812454, + 0.007292131893336773, + 0.007873360998928547, + 0.008103080093860626, + 0.01188061386346817, + 0.005195509176701307, + 0.003733244491741061, + 0.0059699914418160915, + 0.00647300761193037, + 0.0030203936621546745, + 0.003263098420575261, + 0.004755509085953236, + 0.004558435641229153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.11292989552021027 + } + ], + "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/0000131/group_1/index_tensor([4117, 4117])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_tensor([4117, 4117])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..63de44ff901b3fdcc0be337c5cd529caca75783f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_tensor([4117, 4117])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot, shot from above, begins in a medium shot and slowly advances as the action unfolds. Against a deep, nearly black void, countless golden points of light, radiating incandescent energy, appear out of thin air, like tiny stars. Pulled by an invisible force, they rapidly converge into a rapidly rotating counterclockwise vortex of energy, each particle trailing a dazzling trail of light. As the camera zooms in, this luminous vortex violently contracts inward, its particles colliding and merging to become increasingly dazzling, their light transitioning from brilliant white to a deep amber. At the vortex's core, the outline of a solid, powerful fist begins to coalesce. The fingers are first defined, then swiftly and decisively curled inward and clenched, ultimately forming a fist as if forged from molten gold. The fist's muscles are sharply defined, its knuckles sharp with tension, and visible energy lines flow across its smooth, metallic surface. The last few stray light particles were completely absorbed into the fist, and the entire form completely solidified, hovering steadily in the center of the picture, emitting an awe-inspiring glow. The entire scene was filled with an epic and majestic atmosphere of power transitioning from chaos to order.", + "seed": 11488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_4117_seed_11488_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": -0.12670002039521933, + "vlm_logprob": -0.12670002039521933, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.12670002039521933, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015388213098049164 + ], + "opposite_logprob_tokens": [ + -4.265388011932373 + ], + "logprob_reward": -0.015388213098049164, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015388213098049164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15547941625118256, + 0.1158360168337822, + 0.0521063432097435, + 0.0845421850681305, + 0.1357426941394806, + 0.08774811774492264, + 0.15674500167369843, + 0.0376647487282753, + 0.07311283051967621, + 0.1010226458311081 + ], + "credit_map": [ + 0.01825450174510479, + 0.017447203397750854, + 0.04105774685740471, + 0.017974484711885452, + 0.012287124991416931, + 0.017234882339835167, + 0.013934310525655746, + 0.017289167270064354, + 0.031347595155239105, + 0.007476882543414831, + 0.026548556983470917, + 0.010928338393568993, + 0.009071326814591885, + 0.010565083473920822, + 0.008473164401948452, + 0.011425075121223927, + 0.012919934466481209, + 0.013368384912610054, + 0.005558292381465435, + 0.0036058551631867886, + 0.004698770120739937, + 0.004103631712496281, + 0.003393153427168727, + 0.004458324052393436, + 0.009698556736111641, + 0.004652935080230236, + 0.023429926484823227, + 0.004715993069112301, + 0.00416148453950882, + 0.01501164585351944, + 0.006216088309884071, + 0.01665556989610195, + 0.03279659524559975, + 0.00723106088116765, + 0.02863796427845955, + 0.010667451657354832, + 0.008574013598263264, + 0.012703755870461464, + 0.008447225205600262, + 0.026684632524847984, + 0.01406791340559721, + 0.006288478616625071, + 0.022488165646791458, + 0.008030679076910019, + 0.005126665346324444, + 0.015528477728366852, + 0.006484969984740019, + 0.009732767008244991, + 0.019561951979994774, + 0.007252872455865145, + 0.030085988342761993, + 0.01114426739513874, + 0.010026532225310802, + 0.038222748786211014, + 0.010493648238480091, + 0.02995699644088745, + 0.003625475335866213, + 0.001989250537008047, + 0.00910712219774723, + 0.002719153882935643, + 0.0019277437822893262, + 0.010547093115746975, + 0.002078873571008444, + 0.005670041311532259, + 0.004852036014199257, + 0.0015582744963467121, + 0.020054636523127556, + 0.0029113232158124447, + 0.0019494727021083236, + 0.02940843626856804, + 0.0020169883500784636, + 0.010361663065850735, + 0.013162230141460896, + 0.0052650608122348785, + 0.01815156079828739, + 0.008431094698607922, + 0.00761757604777813, + 0.02526325359940529, + 0.008488129824399948, + 0.014643735252320766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.015388213098049164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373394533991814 + ], + "opposite_logprob_tokens": [ + -3.40873384475708 + ], + "logprob_reward": -0.03373394533991814, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373394533991814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06819790601730347, + 0.028334710747003555, + 0.17848087847232819, + 0.36925509572029114, + 0.034218478947877884, + 0.04960589483380318, + 0.046074531972408295, + 0.03572313115000725, + 0.152066171169281, + 0.03804317116737366 + ], + "credit_map": [ + 0.007922520861029625, + 0.016333624720573425, + 0.010863627307116985, + 0.005887366831302643, + 0.00776958791539073, + 0.006136666983366013, + 0.00698343338444829, + 0.006301075220108032, + 0.003901987336575985, + 0.003925503697246313, + 0.0026485600974410772, + 0.0025742799043655396, + 0.007687113713473082, + 0.002546624979004264, + 0.0025014341808855534, + 0.002549207303673029, + 0.049982115626335144, + 0.04807962477207184, + 0.021411366760730743, + 0.01613517478108406, + 0.013566670008003712, + 0.008902034722268581, + 0.009918135590851307, + 0.010485760867595673, + 0.03775700181722641, + 0.014691467396914959, + 0.13993291556835175, + 0.015033518895506859, + 0.01090226136147976, + 0.035731248557567596, + 0.013726337812840939, + 0.10148035734891891, + 0.0037819468416273594, + 0.0019702021963894367, + 0.008060247637331486, + 0.003267491003498435, + 0.0024584548082202673, + 0.00924456026405096, + 0.0026727174408733845, + 0.0027628587558865547, + 0.007648887578397989, + 0.00306883966550231, + 0.010343940928578377, + 0.0045478614047169685, + 0.003147280542179942, + 0.009322967380285263, + 0.004229424986988306, + 0.007296697236597538, + 0.0062697883695364, + 0.002722625620663166, + 0.008698291145265102, + 0.00435256352648139, + 0.003395947627723217, + 0.008336909115314484, + 0.004314401652663946, + 0.007984006777405739, + 0.004339877981692553, + 0.0015605902299284935, + 0.006987459491938353, + 0.0030699726194143295, + 0.001951434533111751, + 0.007964454591274261, + 0.0036260976921766996, + 0.00622324226424098, + 0.009911528788506985, + 0.003771349089220166, + 0.037040065973997116, + 0.007180983200669289, + 0.005158708430826664, + 0.059271782636642456, + 0.005176009610295296, + 0.024555759504437447, + 0.003969813231378794, + 0.0035034799948334694, + 0.0054115294478833675, + 0.004427733365446329, + 0.003582642413675785, + 0.008232424035668373, + 0.004222968593239784, + 0.004692580085247755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.03373394533991814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742806553840637 + ], + "opposite_logprob_tokens": [ + -0.9742806553840637 + ], + "logprob_reward": -0.4742806553840637, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742806553840637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17755940556526184, + 0.051394231617450714, + 0.08331644535064697, + 0.20665626227855682, + 0.1329907327890396, + 0.13846556842327118, + 0.09247217327356339, + 0.032542355358600616, + 0.054226431995630264, + 0.030376406386494637 + ], + "credit_map": [ + 0.018345627933740616, + 0.043185632675886154, + 0.022540446370840073, + 0.012868239544332027, + 0.014982355758547783, + 0.016740282997488976, + 0.011141804046928883, + 0.03775501251220703, + 0.008657898753881454, + 0.005759887397289276, + 0.007918361574411392, + 0.004772850777953863, + 0.005212990567088127, + 0.006158162839710712, + 0.005286149214953184, + 0.007627931423485279, + 0.011743959039449692, + 0.015658801421523094, + 0.014696027152240276, + 0.00681298179551959, + 0.00738898292183876, + 0.00944216176867485, + 0.006567010190337896, + 0.011006518267095089, + 0.021603593602776527, + 0.011068133637309074, + 0.06606931984424591, + 0.011416452936828136, + 0.011983776465058327, + 0.030934082344174385, + 0.011809459887444973, + 0.04177144169807434, + 0.01516225840896368, + 0.005129940342158079, + 0.03550958260893822, + 0.0073537095449864864, + 0.005649592261761427, + 0.04470594599843025, + 0.006975116673856974, + 0.012504600919783115, + 0.02046039141714573, + 0.010409602895379066, + 0.027832478284835815, + 0.01576399616897106, + 0.011088641360402107, + 0.022147690877318382, + 0.012436128221452236, + 0.018326638266444206, + 0.020080721005797386, + 0.003607225138694048, + 0.021157432347536087, + 0.00641469145193696, + 0.004384081810712814, + 0.017518341541290283, + 0.00591887766495347, + 0.013390796259045601, + 0.004757580813020468, + 0.0013283590087667108, + 0.004827406723052263, + 0.002382473321631551, + 0.001563111669383943, + 0.00684727355837822, + 0.002016016747802496, + 0.008820132352411747, + 0.006450657732784748, + 0.0020965624134987593, + 0.007353551685810089, + 0.004311655648052692, + 0.0033883871510624886, + 0.011326517909765244, + 0.0045410324819386005, + 0.014758069068193436, + 0.003185619367286563, + 0.00133059686049819, + 0.006811311934143305, + 0.0019837631843984127, + 0.0015696077607572079, + 0.009314129129052162, + 0.0022907783277332783, + 0.0038905979599803686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.4742806553840637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787074290215969 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.009787074290215969, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787074290215969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05074523761868477, + 0.03260214999318123, + 0.09935586899518967, + 0.304836630821228, + 0.10145881026983261, + 0.16828526556491852, + 0.03392310068011284, + 0.03377850726246834, + 0.13861466944217682, + 0.03639974445104599 + ], + "credit_map": [ + 0.006145249120891094, + 0.005898410454392433, + 0.010058043524622917, + 0.005581180099397898, + 0.0061858706176280975, + 0.005219274666160345, + 0.005926608107984066, + 0.005730599630624056, + 0.006150370929390192, + 0.003102036425843835, + 0.005174495745450258, + 0.002529179211705923, + 0.004614421632140875, + 0.0042127687484025955, + 0.00237873662263155, + 0.0044401418417692184, + 0.030067089945077896, + 0.028553681448101997, + 0.011375716887414455, + 0.006893553771078587, + 0.007820738479495049, + 0.004890899173915386, + 0.004571995232254267, + 0.005182191263884306, + 0.030894238501787186, + 0.011988332495093346, + 0.11789383739233017, + 0.01148706953972578, + 0.0081391716375947, + 0.02474747784435749, + 0.010027489624917507, + 0.08965901285409927, + 0.006634991150349379, + 0.0027696145698428154, + 0.019669679924845695, + 0.0033259179908782244, + 0.003940402530133724, + 0.04021959751844406, + 0.004504179581999779, + 0.02039443701505661, + 0.028520580381155014, + 0.01103805098682642, + 0.038379404693841934, + 0.014751207083463669, + 0.010400705970823765, + 0.028017710894346237, + 0.012838170863687992, + 0.024339446797966957, + 0.0062780557200312614, + 0.0022290851920843124, + 0.00593092106282711, + 0.00339065445587039, + 0.002516708569601178, + 0.004237363580614328, + 0.00316915987059474, + 0.0061711510643363, + 0.004706770647317171, + 0.0019585269037634134, + 0.0057825432159006596, + 0.0036538965068757534, + 0.0024597919546067715, + 0.006613600999116898, + 0.003614801447838545, + 0.00498857581987977, + 0.010484261438250542, + 0.004621138330549002, + 0.034038372337818146, + 0.008767994120717049, + 0.005442134570330381, + 0.04975779727101326, + 0.006595100276172161, + 0.018907863646745682, + 0.004475567489862442, + 0.003013428533449769, + 0.006160543765872717, + 0.00434749573469162, + 0.0034978752955794334, + 0.005684330593794584, + 0.004847168456763029, + 0.004373336676508188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.009787074290215969 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031021386384964 + ], + "opposite_logprob_tokens": [ + -2.3503103256225586 + ], + "logprob_reward": -0.10031021386384964, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031021386384964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04230782762169838, + 0.018383869901299477, + 0.045203231275081635, + 0.2331567108631134, + 0.022022875025868416, + 0.10492728650569916, + 0.0962773934006691, + 0.08833682537078857, + 0.27423423528671265, + 0.07514975965023041 + ], + "credit_map": [ + 0.005219615995883942, + 0.006332290358841419, + 0.007904618978500366, + 0.005366906989365816, + 0.004079799633473158, + 0.004969009198248386, + 0.0037130750715732574, + 0.00472251744940877, + 0.004636114928871393, + 0.002269911114126444, + 0.001647433266043663, + 0.0009504829067736864, + 0.0048482888378202915, + 0.0016619726084172726, + 0.0011409996077418327, + 0.0012286684941500425, + 0.005757474340498447, + 0.008245532400906086, + 0.005911638494580984, + 0.006719287950545549, + 0.004784089513123035, + 0.004522206727415323, + 0.004969256930053234, + 0.004293748643249273, + 0.023787638172507286, + 0.009318106807768345, + 0.08974272012710571, + 0.008625165559351444, + 0.005962342955172062, + 0.020664680749177933, + 0.00813268031924963, + 0.06692337989807129, + 0.0029196874238550663, + 0.0024316415656358004, + 0.0036272509023547173, + 0.0026352631393820047, + 0.002473779721185565, + 0.002031812211498618, + 0.0026006975676864386, + 0.0033027452882379293, + 0.014650934375822544, + 0.008136948570609093, + 0.018014052882790565, + 0.011242936365306377, + 0.007768266834318638, + 0.02280852384865284, + 0.010904749855399132, + 0.01140088215470314, + 0.013517376035451889, + 0.004583448637276888, + 0.020204296335577965, + 0.007129302714020014, + 0.0056706746108829975, + 0.021219128742814064, + 0.0065218498930335045, + 0.01743132434785366, + 0.00997909065335989, + 0.00685378210619092, + 0.009989652782678604, + 0.008123045787215233, + 0.007888831198215485, + 0.018736133351922035, + 0.012518229894340038, + 0.01424806285649538, + 0.017890680581331253, + 0.0063956682570278645, + 0.07052401453256607, + 0.011561141349375248, + 0.008258624002337456, + 0.11034324765205383, + 0.008980432525277138, + 0.04028042405843735, + 0.008059980347752571, + 0.0067566800862550735, + 0.01053653284907341, + 0.009506875649094582, + 0.006704385858029127, + 0.013043621554970741, + 0.009595522657036781, + 0.010946161113679409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 5 + ], + "selected_reward": -0.10031021386384964 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from above, begins in a medium shot and slowly advances as the action unfolds. Against a deep, nearly black void, countless golden points of light, radiating incandescent energy, appear out of thin air, like tiny stars. Pulled by an invisible force, they rapidly converge into a rapidly rotating counterclockwise vortex of energy, each particle trailing a dazzling trail of light. As the camera zooms in, this luminous vortex violently contracts inward, its particles colliding and merging to become increasingly dazzling, their light transitioning from brilliant white to a deep amber. At the vortex's core, the outline of a solid, powerful fist begins to coalesce. The fingers are first defined, then swiftly and decisively curled inward and clenched, ultimately forming a fist as if forged from molten gold. The fist's muscles are sharply defined, its knuckles sharp with tension, and visible energy lines flow across its smooth, metallic surface. The last few stray light particles were completely absorbed into the fist, and the entire form completely solidified, hovering steadily in the center of the picture, emitting an awe-inspiring glow. The entire scene was filled with an epic and majestic atmosphere of power transitioning from chaos to order.", + "seed": 11489, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_4117_seed_11489_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": -0.186954673146829, + "vlm_logprob": -0.186954673146829, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.186954673146829, + "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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0012191252317279577 + ], + "opposite_logprob_tokens": [ + -7.251219272613525 + ], + "logprob_reward": -0.0012191252317279577, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0012191252317279577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2350425124168396, + 0.1340450793504715, + 0.0705028623342514, + 0.020625246688723564, + 0.06250479072332382, + 0.032404642552137375, + 0.03875408694148064, + 0.1028684601187706, + 0.26327821612358093, + 0.03997404873371124 + ], + "credit_map": [ + 0.03296084329485893, + 0.011858856305480003, + 0.02140674740076065, + 0.01701459474861622, + 0.014379427768290043, + 0.04953385517001152, + 0.01921885833144188, + 0.06866933405399323, + 0.014764017425477505, + 0.008042605593800545, + 0.011394153349101543, + 0.01009217370301485, + 0.010653818026185036, + 0.027257002890110016, + 0.00956685934215784, + 0.042274441570043564, + 0.008310945704579353, + 0.004458072129637003, + 0.006293970625847578, + 0.004268513526767492, + 0.004391304217278957, + 0.016641465947031975, + 0.008153927512466908, + 0.01798466593027115, + 0.0026825929526239634, + 0.001925442717038095, + 0.0028071319684386253, + 0.0029450280126184225, + 0.002318752696737647, + 0.0031419135630130768, + 0.002489817328751087, + 0.0023145671002566814, + 0.009232930839061737, + 0.0073761907406151295, + 0.008323400281369686, + 0.008647322654724121, + 0.004794462118297815, + 0.006613662466406822, + 0.007827009074389935, + 0.009689809754490852, + 0.004102629609405994, + 0.0033007757738232613, + 0.003943454008549452, + 0.005809498019516468, + 0.0031061721965670586, + 0.003791628871113062, + 0.004003697540611029, + 0.004346788860857487, + 0.004082693252712488, + 0.003495205193758011, + 0.004725377541035414, + 0.004089681897312403, + 0.005299697630107403, + 0.0049890922382473946, + 0.004383671563118696, + 0.007688669487833977, + 0.010804804973304272, + 0.005646037869155407, + 0.022012192755937576, + 0.008461646735668182, + 0.008063195273280144, + 0.015924576669931412, + 0.009981155395507812, + 0.021974852308630943, + 0.013965992256999016, + 0.006793585605919361, + 0.08704697340726852, + 0.007988092489540577, + 0.010050172917544842, + 0.10745421797037125, + 0.009050815366208553, + 0.020928382873535156, + 0.005056430120021105, + 0.0018711561569944024, + 0.009343685582280159, + 0.002189334249123931, + 0.0023981528356671333, + 0.007127939257770777, + 0.0023562998976558447, + 0.009631050750613213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0012191252317279577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215397648513317 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215397648513317, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215397648513317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15880176424980164, + 0.18872003257274628, + 0.09732852131128311, + 0.0827179029583931, + 0.11129362136125565, + 0.044307734817266464, + 0.028806136921048164, + 0.047578878700733185, + 0.09857521206140518, + 0.14187023043632507 + ], + "credit_map": [ + 0.02346860058605671, + 0.009329153224825859, + 0.013673975132405758, + 0.00833082851022482, + 0.013175922445952892, + 0.03279461711645126, + 0.00898713432252407, + 0.04904153570532799, + 0.02055250108242035, + 0.009691521525382996, + 0.016722645610570908, + 0.012498734518885612, + 0.014825568534433842, + 0.039680637419223785, + 0.013663285411894321, + 0.061085134744644165, + 0.007042956538498402, + 0.00712336553260684, + 0.005984880495816469, + 0.007716069929301739, + 0.008637727238237858, + 0.022961188107728958, + 0.013863163068890572, + 0.0239991694688797, + 0.010132590308785439, + 0.008450997062027454, + 0.009653604589402676, + 0.012694782577455044, + 0.009875222109258175, + 0.00955910049378872, + 0.012084214016795158, + 0.010267387144267559, + 0.015868308022618294, + 0.013351570814847946, + 0.013315913267433643, + 0.013906341977417469, + 0.011025039479136467, + 0.01223510317504406, + 0.013077558018267155, + 0.01851377822458744, + 0.005570479668676853, + 0.005336647387593985, + 0.006270323880016804, + 0.0064997063018381596, + 0.006281143054366112, + 0.00519189890474081, + 0.005115342326462269, + 0.0040421937592327595, + 0.0045449691824615, + 0.003417686326429248, + 0.004450822249054909, + 0.002290771109983325, + 0.002282399218529463, + 0.0038944175466895103, + 0.002580533269792795, + 0.0053445384837687016, + 0.005225352477282286, + 0.0038024757523089647, + 0.008043802343308926, + 0.005076207686215639, + 0.004075472708791494, + 0.007649372331798077, + 0.006028445437550545, + 0.007677748799324036, + 0.007675448432564735, + 0.002828100463375449, + 0.029942799359560013, + 0.00437379302456975, + 0.0039755115285515785, + 0.03490499034523964, + 0.004501771181821823, + 0.010372799821197987, + 0.00861193798482418, + 0.002641799859702587, + 0.05464300140738487, + 0.0032451190054416656, + 0.0032066546846181154, + 0.05624983459711075, + 0.0028613682370632887, + 0.010410505346953869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.014215397648513317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898404449224472 + ], + "opposite_logprob_tokens": [ + -2.578984022140503 + ], + "logprob_reward": -0.07898404449224472, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898404449224472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12280959635972977, + 0.06710730493068695, + 0.06137192249298096, + 0.040599215775728226, + 0.044449787586927414, + 0.036123502999544144, + 0.051391780376434326, + 0.21297430992126465, + 0.3357613682746887, + 0.027411233633756638 + ], + "credit_map": [ + 0.016649505123496056, + 0.008307025767862797, + 0.013066992163658142, + 0.009185794740915298, + 0.00819175224751234, + 0.025933722034096718, + 0.007798082195222378, + 0.03367672115564346, + 0.005708273034542799, + 0.005608333274722099, + 0.0066228327341377735, + 0.005642435047775507, + 0.00821459200233221, + 0.011302380822598934, + 0.00673656165599823, + 0.017271891236305237, + 0.005108549725264311, + 0.0075194817036390305, + 0.006744137033820152, + 0.006920047104358673, + 0.008223417215049267, + 0.006426330655813217, + 0.013505887240171432, + 0.006924070883542299, + 0.003959165420383215, + 0.0024170067626982927, + 0.0105137312784791, + 0.005234828218817711, + 0.003823331091552973, + 0.004264721646904945, + 0.0046448493376374245, + 0.00574158038944006, + 0.005900033749639988, + 0.004581410437822342, + 0.0047193365171551704, + 0.007958246394991875, + 0.003952070139348507, + 0.005049447529017925, + 0.005832982715219259, + 0.006456260569393635, + 0.004675946664065123, + 0.00458024675026536, + 0.003922686446458101, + 0.005705923307687044, + 0.004538302775472403, + 0.004819617606699467, + 0.004767447244375944, + 0.0031133315060287714, + 0.0049270447343587875, + 0.004481784999370575, + 0.006580033805221319, + 0.005145900882780552, + 0.008811749517917633, + 0.0080520985648036, + 0.0072214058600366116, + 0.00617176340892911, + 0.024836929515004158, + 0.012356129474937916, + 0.04336414858698845, + 0.014293120242655277, + 0.015656858682632446, + 0.03581425920128822, + 0.017495740205049515, + 0.04915712773799896, + 0.02108689956367016, + 0.012100656516849995, + 0.09631451219320297, + 0.012633632868528366, + 0.01953594572842121, + 0.12335240840911865, + 0.016804736107587814, + 0.03393260017037392, + 0.001581269083544612, + 0.0005786017281934619, + 0.010491752997040749, + 0.0006756081711500883, + 0.0006461372831836343, + 0.010975227691233158, + 0.0007705656462348998, + 0.0016920698108151555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.07898404449224472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198003336787224 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014198003336787224, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198003336787224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04428378492593765, + 0.152961865067482, + 0.10422936826944351, + 0.031516823917627335, + 0.15527573227882385, + 0.015669263899326324, + 0.029407164081931114, + 0.11078985780477524, + 0.2246638834476471, + 0.13120220601558685 + ], + "credit_map": [ + 0.00837025512009859, + 0.0035261353477835655, + 0.0069816820323467255, + 0.004839412402361631, + 0.003277041018009186, + 0.005820723716169596, + 0.004386759363114834, + 0.007081774063408375, + 0.016885904595255852, + 0.007928159087896347, + 0.011527009308338165, + 0.010408594273030758, + 0.011898859404027462, + 0.03245720639824867, + 0.011600767262279987, + 0.050255369395017624, + 0.016519945114850998, + 0.010865353047847748, + 0.014934043399989605, + 0.010700526647269726, + 0.011518244631588459, + 0.011504321359097958, + 0.0174945667386055, + 0.010692369192838669, + 0.0035830228589475155, + 0.0032332774717360735, + 0.005186773370951414, + 0.004845284856855869, + 0.003803375642746687, + 0.003947390243411064, + 0.003700434695929289, + 0.0032172666396945715, + 0.02022206038236618, + 0.01946406438946724, + 0.01966199092566967, + 0.0223525483161211, + 0.012094377540051937, + 0.014969303272664547, + 0.021848278120160103, + 0.024663107469677925, + 0.0016083481023088098, + 0.002209268044680357, + 0.0016783589962869883, + 0.0024329470470547676, + 0.0016852085245773196, + 0.0019164029508829117, + 0.0026424950920045376, + 0.0014962333953008056, + 0.004042441491037607, + 0.0024642881471663713, + 0.003868918167427182, + 0.003230820409953594, + 0.0028965992387384176, + 0.004185003228485584, + 0.0037905965000391006, + 0.0049284957349300385, + 0.01473624724894762, + 0.0059582446701824665, + 0.02080428972840309, + 0.007267974317073822, + 0.005398668348789215, + 0.017036834731698036, + 0.00976432766765356, + 0.02982327528297901, + 0.01456368062645197, + 0.004454534966498613, + 0.08059865981340408, + 0.005875336471945047, + 0.0061841486021876335, + 0.08821132779121399, + 0.006071672774851322, + 0.018704529851675034, + 0.00942215882241726, + 0.0027015497907996178, + 0.046148814260959625, + 0.003459751373156905, + 0.0036242364440113306, + 0.04748423397541046, + 0.004182273987680674, + 0.014179188758134842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.014198003336787224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8261567950248718 + ], + "opposite_logprob_tokens": [ + -0.5761567950248718 + ], + "logprob_reward": -0.8261567950248718, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261567950248718, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.020772729068994522, + 0.003125028219074011, + 0.004873440135270357, + 0.004750561900436878, + 0.021206330507993698, + 0.013498587533831596, + 0.008307620882987976, + 0.06390400230884552, + 0.0549464225769043, + 0.8046152591705322 + ], + "credit_map": [ + 0.0023706855718046427, + 0.0016237461240962148, + 0.0036899500992149115, + 0.0017455011839047074, + 0.001619336660951376, + 0.0037332132924348116, + 0.0014597033150494099, + 0.004530593287199736, + 0.00026940956013277173, + 0.0003341049887239933, + 0.00024999314337037504, + 0.00027499525458551943, + 0.0005043279379606247, + 0.0005478864768519998, + 0.00039111115620471537, + 0.0005532000213861465, + 0.0005080465343780816, + 0.000416799244703725, + 0.0004853339632973075, + 0.0003759291721507907, + 0.0004581657121889293, + 0.0009520627208985388, + 0.0005543885636143386, + 0.0011227146023884416, + 0.0009707330609671772, + 0.0003145052760373801, + 0.0008244704222306609, + 0.0003505535132717341, + 0.0003414643870200962, + 0.0008432588074356318, + 0.0004233524377923459, + 0.0006822244613431394, + 0.002715901704505086, + 0.002161748008802533, + 0.002422775374725461, + 0.0031437480356544256, + 0.0017795668682083488, + 0.0025420833844691515, + 0.002585015492513776, + 0.0038554940838366747, + 0.0015926832566037774, + 0.0013562565436586738, + 0.0016076673055067658, + 0.0021642495412379503, + 0.0015649169217795134, + 0.0019783396273851395, + 0.001770111615769565, + 0.0014643640024587512, + 0.0008565631578676403, + 0.0007910261629149318, + 0.0010058567859232426, + 0.0008713373681530356, + 0.0010841261828318238, + 0.0013134567998349667, + 0.0012531154789030552, + 0.0011321400525048375, + 0.008506986312568188, + 0.0034271746408194304, + 0.01126904971897602, + 0.004391027148813009, + 0.003252242226153612, + 0.010312494821846485, + 0.005484725348651409, + 0.01726030930876732, + 0.00394415482878685, + 0.001450896612368524, + 0.01785069890320301, + 0.0021108018700033426, + 0.001810807385481894, + 0.02004031278192997, + 0.0018777836812660098, + 0.005860975477844477, + 0.04102068766951561, + 0.010554403997957706, + 0.32409170269966125, + 0.013399435207247734, + 0.011388485319912434, + 0.3465927839279175, + 0.009704080410301685, + 0.0478636771440506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.8261567950248718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "shot, shot from above, begins in a medium shot and slowly advances as the action unfolds. Against a deep, nearly black void, countless golden points of light, radiating incandescent energy, appear out of thin air, like tiny stars. Pulled by an invisible force, they rapidly converge into a rapidly rotating counterclockwise vortex of energy, each particle trailing a dazzling trail of light. As the camera zooms in, this luminous vortex violently contracts inward, its particles colliding and merging to become increasingly dazzling, their light transitioning from brilliant white to a deep amber. At the vortex's core, the outline of a solid, powerful fist begins to coalesce. The fingers are first defined, then swiftly and decisively curled inward and clenched, ultimately forming a fist as if forged from molten gold. The fist's muscles are sharply defined, its knuckles sharp with tension, and visible energy lines flow across its smooth, metallic surface. The last few stray light particles were completely absorbed into the fist, and the entire form completely solidified, hovering steadily in the center of the picture, emitting an awe-inspiring glow. The entire scene was filled with an epic and majestic atmosphere of power transitioning from chaos to order.", + "seed": 11490, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_4117_seed_11490_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": -0.12855409374460577, + "vlm_logprob": -0.12855409374460577, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.12855409374460577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004244011361151934 + ], + "opposite_logprob_tokens": [ + -5.754243850708008 + ], + "logprob_reward": -0.004244011361151934, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004244011361151934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08983774483203888, + 0.06842899322509766, + 0.027230920270085335, + 0.10105293989181519, + 0.027588894590735435, + 0.18657706677913666, + 0.03292281553149223, + 0.14882540702819824, + 0.27534425258636475, + 0.04219092056155205 + ], + "credit_map": [ + 0.009873045608401299, + 0.008855994790792465, + 0.0111372796818614, + 0.007666905876249075, + 0.020676294341683388, + 0.011710151098668575, + 0.009658859111368656, + 0.01025922130793333, + 0.0054319133050739765, + 0.005063391290605068, + 0.006362035870552063, + 0.008201955817639828, + 0.007916166447103024, + 0.013788307085633278, + 0.009184487164020538, + 0.012480737641453743, + 0.0047737048007547855, + 0.0023939087986946106, + 0.002304498804733157, + 0.0027470234781503677, + 0.003719854634255171, + 0.0029024891555309296, + 0.004966575652360916, + 0.0034228640142828226, + 0.0076330252923071384, + 0.006495433859527111, + 0.03385107219219208, + 0.004535285755991936, + 0.0061194682493805885, + 0.027335701510310173, + 0.0031038157176226377, + 0.011979131028056145, + 0.002641703700646758, + 0.003089914796873927, + 0.00436040386557579, + 0.0031783708836883307, + 0.0032257495913654566, + 0.004749420564621687, + 0.002820002380758524, + 0.0035233290400356054, + 0.036421358585357666, + 0.009586024098098278, + 0.015072822570800781, + 0.015007827430963516, + 0.012230385094881058, + 0.03330518677830696, + 0.011840315535664558, + 0.05311316251754761, + 0.003553494345396757, + 0.0030082270968705416, + 0.0047433506697416306, + 0.00558478245511651, + 0.003663343843072653, + 0.004854890052229166, + 0.0034788416232913733, + 0.004035884514451027, + 0.02005571499466896, + 0.011656267568469048, + 0.02888689562678337, + 0.02193361334502697, + 0.011522850953042507, + 0.020957693457603455, + 0.017127079889178276, + 0.01668529585003853, + 0.03131799399852753, + 0.007689040154218674, + 0.07015449553728104, + 0.00905148871243, + 0.007331477012485266, + 0.031793463975191116, + 0.011540662497282028, + 0.10646563768386841, + 0.01057595293968916, + 0.002833797363564372, + 0.0032336777076125145, + 0.002694264752790332, + 0.003584581194445491, + 0.005013100802898407, + 0.003608984174206853, + 0.010646559298038483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.004244011361151934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016096990555524826 + ], + "opposite_logprob_tokens": [ + -4.141097068786621 + ], + "logprob_reward": -0.016096990555524826, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016096990555524826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11256242543458939, + 0.08685016632080078, + 0.0842893198132515, + 0.19607287645339966, + 0.021038688719272614, + 0.10165831446647644, + 0.0825909897685051, + 0.06905614584684372, + 0.15700437128543854, + 0.08887666463851929 + ], + "credit_map": [ + 0.016024149954319, + 0.009461828507483006, + 0.016760841012001038, + 0.008212433196604252, + 0.017650539055466652, + 0.015187541022896767, + 0.013341984711587429, + 0.01592310518026352, + 0.005843574646860361, + 0.006938562728464603, + 0.006411004811525345, + 0.008143694140017033, + 0.013262110762298107, + 0.018850743770599365, + 0.010117964819073677, + 0.017282508313655853, + 0.008282898925244808, + 0.008266236633062363, + 0.010473824106156826, + 0.008762548677623272, + 0.012623734772205353, + 0.008868204429745674, + 0.015624784864485264, + 0.011387082748115063, + 0.014052948914468288, + 0.013734154403209686, + 0.06458491086959839, + 0.008771486580371857, + 0.011807071045041084, + 0.04747265577316284, + 0.007354171946644783, + 0.028295477852225304, + 0.0029378500767052174, + 0.002720027696341276, + 0.0024761203676462173, + 0.002861177083104849, + 0.0020002357196062803, + 0.0029324444476515055, + 0.001985265174880624, + 0.003125569550320506, + 0.012543418444693089, + 0.010040942579507828, + 0.008396616205573082, + 0.017452698200941086, + 0.015707623213529587, + 0.014681199565529823, + 0.008172301575541496, + 0.014663506299257278, + 0.009559600614011288, + 0.006800197064876556, + 0.016022304072976112, + 0.012252802960574627, + 0.0068125915713608265, + 0.01047058217227459, + 0.007486279588192701, + 0.013186630792915821, + 0.01151048019528389, + 0.00350134982727468, + 0.010717133060097694, + 0.005199312698096037, + 0.004222639370709658, + 0.01280183158814907, + 0.004278597887605429, + 0.016824796795845032, + 0.017678001895546913, + 0.004599674139171839, + 0.038683097809553146, + 0.005398084409534931, + 0.005627925042062998, + 0.017988650128245354, + 0.008441089652478695, + 0.05858785659074783, + 0.025156499817967415, + 0.004790271166712046, + 0.008542370051145554, + 0.004383541643619537, + 0.006446989253163338, + 0.011733399704098701, + 0.00548900943249464, + 0.02233458310365677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.016096990555524826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306282848119736 + ], + "opposite_logprob_tokens": [ + -3.168062925338745 + ], + "logprob_reward": -0.04306282848119736, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306282848119736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02780763991177082, + 0.02783181704580784, + 0.014745768159627914, + 0.43511199951171875, + 0.025011690333485603, + 0.07691348344087601, + 0.10685545206069946, + 0.15042424201965332, + 0.11411374062299728, + 0.02118421532213688 + ], + "credit_map": [ + 0.004208078607916832, + 0.001748836599290371, + 0.004323834087699652, + 0.0023218032438308, + 0.004070993512868881, + 0.004064320120960474, + 0.0034077237360179424, + 0.0036620493046939373, + 0.0021221665665507317, + 0.0018543392652645707, + 0.0022371422965079546, + 0.002386180916801095, + 0.0033110445365309715, + 0.006622595712542534, + 0.0032153308857232332, + 0.006083013489842415, + 0.0020079032983630896, + 0.0019005017820745707, + 0.0024736050982028246, + 0.0016524160746484995, + 0.002051473595201969, + 0.00162024749442935, + 0.0015209988923743367, + 0.0015186212258413434, + 0.03201865777373314, + 0.027646973729133606, + 0.14677774906158447, + 0.018787125125527382, + 0.025131627917289734, + 0.11438942700624466, + 0.013410953804850578, + 0.05694945156574249, + 0.0026081360410898924, + 0.002978083910420537, + 0.003353265579789877, + 0.0038182116113603115, + 0.0023788204416632652, + 0.003134741447865963, + 0.0025686833541840315, + 0.0041717467829585075, + 0.014831975102424622, + 0.004307125229388475, + 0.006514235865324736, + 0.00468352809548378, + 0.0046267020516097546, + 0.014549467712640762, + 0.004536708816885948, + 0.022863740101456642, + 0.012452623806893826, + 0.009090294130146503, + 0.019814729690551758, + 0.01877952180802822, + 0.009305967949330807, + 0.013346603140234947, + 0.009853924624621868, + 0.014211773872375488, + 0.01953948475420475, + 0.010571382008492947, + 0.031454455107450485, + 0.02101532742381096, + 0.010863907635211945, + 0.023572437465190887, + 0.01599884405732155, + 0.017408397048711777, + 0.013305912725627422, + 0.0023576777894049883, + 0.02883124351501465, + 0.003392784623429179, + 0.0026765966322273016, + 0.013901909813284874, + 0.0041489508002996445, + 0.04549865424633026, + 0.004753194283694029, + 0.0015304480912163854, + 0.0016042866045609117, + 0.0018951001111418009, + 0.002066944958642125, + 0.002732262248173356, + 0.0018953124526888132, + 0.0047066668048501015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.04306282848119736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06036408245563507, + 0.07488294690847397, + 0.028393473476171494, + 0.15287058055400848, + 0.037463199347257614, + 0.045611295849084854, + 0.07443762570619583, + 0.13160006701946259, + 0.24843688309192657, + 0.1459398865699768 + ], + "credit_map": [ + 0.007566570304334164, + 0.0062292288057506084, + 0.0057893176563084126, + 0.006171385757625103, + 0.011274582706391811, + 0.007632732391357422, + 0.006768549792468548, + 0.008931707590818405, + 0.0054816859774291515, + 0.006446139886975288, + 0.005832028109580278, + 0.008672255091369152, + 0.013250189833343029, + 0.013728916645050049, + 0.01046915166079998, + 0.01100256945937872, + 0.003232735674828291, + 0.00286362087354064, + 0.004411065019667149, + 0.00267878663726151, + 0.0037658943329006433, + 0.003009824315086007, + 0.005864433478564024, + 0.0025671112816780806, + 0.011181577108800411, + 0.009854218922555447, + 0.05142413079738617, + 0.0065490854904055595, + 0.009156947955489159, + 0.04120628163218498, + 0.004963678773492575, + 0.018534649163484573, + 0.002710031345486641, + 0.005040909629315138, + 0.005018523428589106, + 0.006647156085819006, + 0.004432218614965677, + 0.0039345258846879005, + 0.004790078848600388, + 0.0048897480592131615, + 0.008363762870430946, + 0.002730690175667405, + 0.003674300853163004, + 0.0035798363387584686, + 0.003324510296806693, + 0.007920834235846996, + 0.0030187664087861776, + 0.012998590245842934, + 0.009919448755681515, + 0.006621775217354298, + 0.016820352524518967, + 0.006623127032071352, + 0.00664939871057868, + 0.011649844236671925, + 0.0062797050923109055, + 0.009873972274363041, + 0.018319105729460716, + 0.009945237077772617, + 0.025056898593902588, + 0.01792927458882332, + 0.009934311732649803, + 0.017907898873090744, + 0.014764094725251198, + 0.017743239179253578, + 0.02558998391032219, + 0.008197023533284664, + 0.0662020817399025, + 0.009192193858325481, + 0.007242628373205662, + 0.026389876380562782, + 0.012218884192407131, + 0.09340419620275497, + 0.0398898646235466, + 0.006019884720444679, + 0.011916834861040115, + 0.009276455268263817, + 0.012470090761780739, + 0.013784690760076046, + 0.01173448283225298, + 0.04084756225347519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.003195776604115963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761708617210388 + ], + "opposite_logprob_tokens": [ + -0.8261708617210388 + ], + "logprob_reward": -0.5761708617210388, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761708617210388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05920149385929108, + 0.01112403441220522, + 0.020735878497362137, + 0.09826617687940598, + 0.017136873677372932, + 0.0592767708003521, + 0.12945176661014557, + 0.06394718587398529, + 0.09531039744615555, + 0.4455493986606598 + ], + "credit_map": [ + 0.007888487540185452, + 0.004019875545054674, + 0.010109229944646358, + 0.004796287510544062, + 0.007681016344577074, + 0.008172835223376751, + 0.0071333954110741615, + 0.009400371462106705, + 0.000974479888100177, + 0.0008565244497731328, + 0.0012323356932029128, + 0.001281354227103293, + 0.0011946085141971707, + 0.0014077710220590234, + 0.002096539596095681, + 0.002080421196296811, + 0.00414104713127017, + 0.0019931683782488108, + 0.003069527680054307, + 0.0020135934464633465, + 0.0025460757315158844, + 0.002620960585772991, + 0.002185782650485635, + 0.0021657219622284174, + 0.007305762730538845, + 0.006390511058270931, + 0.03249108046293259, + 0.004402169957756996, + 0.005438951309770346, + 0.02269819565117359, + 0.003455623285844922, + 0.01608388125896454, + 0.0015767504228278995, + 0.002173465443775058, + 0.002610862487927079, + 0.0022369269281625748, + 0.0015899485442787409, + 0.0025856373831629753, + 0.0019143258687108755, + 0.002448958344757557, + 0.006910492200404406, + 0.006099121645092964, + 0.004371537361294031, + 0.009305134415626526, + 0.007752938661724329, + 0.00922035425901413, + 0.005221001338213682, + 0.010396191850304604, + 0.01638215035200119, + 0.011617149226367474, + 0.02051028423011303, + 0.02397828735411167, + 0.01521763764321804, + 0.015142299234867096, + 0.012445189990103245, + 0.014158772304654121, + 0.008717902936041355, + 0.005221780855208635, + 0.011228283867239952, + 0.009145467542111874, + 0.00496171647682786, + 0.008131985552608967, + 0.008262638002634048, + 0.008277415297925472, + 0.017075393348932266, + 0.004862940404564142, + 0.011795783415436745, + 0.0042815376073122025, + 0.0045428466983139515, + 0.014309493824839592, + 0.004596248734742403, + 0.033846158534288406, + 0.11786890029907227, + 0.02984778769314289, + 0.031487829983234406, + 0.034955766052007675, + 0.03923681005835533, + 0.04311981424689293, + 0.03418644145131111, + 0.11484605073928833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.5761708617210388 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from above, begins in a medium shot and slowly advances as the action unfolds. Against a deep, nearly black void, countless golden points of light, radiating incandescent energy, appear out of thin air, like tiny stars. Pulled by an invisible force, they rapidly converge into a rapidly rotating counterclockwise vortex of energy, each particle trailing a dazzling trail of light. As the camera zooms in, this luminous vortex violently contracts inward, its particles colliding and merging to become increasingly dazzling, their light transitioning from brilliant white to a deep amber. At the vortex's core, the outline of a solid, powerful fist begins to coalesce. The fingers are first defined, then swiftly and decisively curled inward and clenched, ultimately forming a fist as if forged from molten gold. The fist's muscles are sharply defined, its knuckles sharp with tension, and visible energy lines flow across its smooth, metallic surface. The last few stray light particles were completely absorbed into the fist, and the entire form completely solidified, hovering steadily in the center of the picture, emitting an awe-inspiring glow. The entire scene was filled with an epic and majestic atmosphere of power transitioning from chaos to order.", + "seed": 11491, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_1/index_4117_seed_11491_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": -0.02076502637937665, + "vlm_logprob": -0.02076502637937665, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02076502637937665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005752911325544119 + ], + "opposite_logprob_tokens": [ + -5.255753040313721 + ], + "logprob_reward": -0.005752911325544119, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005752911325544119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2274514138698578, + 0.05143201723694801, + 0.06534451246261597, + 0.054860670119524, + 0.2503523528575897, + 0.03941843658685684, + 0.05698174610733986, + 0.14696699380874634, + 0.03304029256105423, + 0.07415150105953217 + ], + "credit_map": [ + 0.024651063606142998, + 0.024669187143445015, + 0.028392501175403595, + 0.028295036405324936, + 0.02214748226106167, + 0.03146223723888397, + 0.029164062812924385, + 0.03866984695196152, + 0.006738422904163599, + 0.003619225462898612, + 0.00995373073965311, + 0.00548697542399168, + 0.004382610786706209, + 0.007792436517775059, + 0.0042784190736711025, + 0.009180200286209583, + 0.009094751439988613, + 0.006570516619831324, + 0.00850726943463087, + 0.0060724616050720215, + 0.007849368266761303, + 0.009338177740573883, + 0.008109088987112045, + 0.009802872315049171, + 0.0030208337120711803, + 0.006600833497941494, + 0.00500422203913331, + 0.007504857145249844, + 0.010976583696901798, + 0.0061010271310806274, + 0.010577799752354622, + 0.00507451081648469, + 0.0309604499489069, + 0.02393949031829834, + 0.05025988817214966, + 0.01787404902279377, + 0.019696416333317757, + 0.029820749536156654, + 0.03097059763967991, + 0.04683073237538338, + 0.004256266634911299, + 0.0038540754467248917, + 0.005977751687169075, + 0.005298360716551542, + 0.004128963686525822, + 0.0061141797341406345, + 0.005271852947771549, + 0.00451698387041688, + 0.009280495345592499, + 0.005011739209294319, + 0.017215652391314507, + 0.0054674046114087105, + 0.004290180280804634, + 0.007194616831839085, + 0.004243307746946812, + 0.004278352949768305, + 0.04959464445710182, + 0.0052282316610217094, + 0.011644164100289345, + 0.005149747710675001, + 0.006906847003847361, + 0.04705223813652992, + 0.004885515663772821, + 0.016505610197782516, + 0.007247532717883587, + 0.003351356368511915, + 0.0053833965212106705, + 0.0027818125672638416, + 0.0038119147066026926, + 0.0029813572764396667, + 0.003363113384693861, + 0.004119809251278639, + 0.016360213980078697, + 0.008844122290611267, + 0.007881109602749348, + 0.0076544880867004395, + 0.010155348107218742, + 0.005031749606132507, + 0.006270588841289282, + 0.01195387914776802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.005752911325544119 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215397648513317 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215397648513317, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215397648513317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1367369145154953, + 0.13778486847877502, + 0.05091284215450287, + 0.24648869037628174, + 0.036090973764657974, + 0.0515974685549736, + 0.07704012095928192, + 0.1785498410463333, + 0.021534224972128868, + 0.06326405704021454 + ], + "credit_map": [ + 0.013659459538757801, + 0.017026502639055252, + 0.018040450289845467, + 0.01667669229209423, + 0.017948394641280174, + 0.021335303783416748, + 0.01713656820356846, + 0.01491353940218687, + 0.017399705946445465, + 0.007912930101156235, + 0.029703805223107338, + 0.008847178891301155, + 0.010627374984323978, + 0.022527456283569336, + 0.00913462694734335, + 0.03163179010152817, + 0.0072908480651676655, + 0.005240705795586109, + 0.0072380490601062775, + 0.005707117263227701, + 0.004475035239011049, + 0.007354413159191608, + 0.007210528012365103, + 0.006396148353815079, + 0.014983680099248886, + 0.018428513780236244, + 0.020293837413191795, + 0.016968045383691788, + 0.026394259184598923, + 0.07467509061098099, + 0.02712189592421055, + 0.04762338101863861, + 0.003998952452093363, + 0.0035931081511080265, + 0.005682165268808603, + 0.0025731874629855156, + 0.004008107353001833, + 0.0049775997176766396, + 0.004268941935151815, + 0.006988911423832178, + 0.005766148678958416, + 0.004207898396998644, + 0.006774981040507555, + 0.006157486699521542, + 0.005841798614710569, + 0.006385575048625469, + 0.007851175963878632, + 0.008612405508756638, + 0.012496811337769032, + 0.006794345565140247, + 0.013488389551639557, + 0.007197021506726742, + 0.008671076036989689, + 0.010625033639371395, + 0.00713270204141736, + 0.010634740814566612, + 0.04543140158057213, + 0.005716548301279545, + 0.008200166746973991, + 0.005487929098308086, + 0.008133221417665482, + 0.06791732460260391, + 0.005519290920346975, + 0.03214395418763161, + 0.0033916952088475227, + 0.0023310931865125895, + 0.003292227629572153, + 0.0024474814999848604, + 0.0033028461039066315, + 0.0018292733002454042, + 0.001847754931077361, + 0.003091853577643633, + 0.016862092539668083, + 0.006551499478518963, + 0.007058082614094019, + 0.005650834646075964, + 0.00660084979608655, + 0.0044486564584076405, + 0.003862122306600213, + 0.012229920364916325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.014215397648513317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620332732796669 + ], + "opposite_logprob_tokens": [ + -2.812033176422119 + ], + "logprob_reward": -0.0620332732796669, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620332732796669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07910984009504318, + 0.04822885990142822, + 0.32616057991981506, + 0.07048900425434113, + 0.06338908523321152, + 0.04969923943281174, + 0.04761514067649841, + 0.21294252574443817, + 0.055775199085474014, + 0.04659053683280945 + ], + "credit_map": [ + 0.009515863843262196, + 0.010288256220519543, + 0.011377890594303608, + 0.007342054508626461, + 0.009027231484651566, + 0.010348579846322536, + 0.00835532695055008, + 0.012854637578129768, + 0.006363807711750269, + 0.00287733250297606, + 0.007793129421770573, + 0.0035662599839270115, + 0.004326795693486929, + 0.0076684304513037205, + 0.0036398323718458414, + 0.011993270367383957, + 0.017093395814299583, + 0.013438865542411804, + 0.016452081501483917, + 0.01382972951978445, + 0.015137121081352234, + 0.12283919006586075, + 0.013273111544549465, + 0.11409710347652435, + 0.005204719956964254, + 0.003979327157139778, + 0.0038256896659731865, + 0.0038227217737585306, + 0.004779009148478508, + 0.024294894188642502, + 0.0045362296514213085, + 0.0200464129447937, + 0.005490601994097233, + 0.00637363875284791, + 0.009590394794940948, + 0.006700590718537569, + 0.006107177585363388, + 0.008385312743484974, + 0.008652430027723312, + 0.012088934890925884, + 0.005499571096152067, + 0.005271600093692541, + 0.006105391308665276, + 0.007997176609933376, + 0.006241410970687866, + 0.005877834744751453, + 0.0077319275587797165, + 0.004974327050149441, + 0.0060091037303209305, + 0.004668243695050478, + 0.0087747722864151, + 0.0034053476992994547, + 0.003952928353101015, + 0.009571610949933529, + 0.004175509326159954, + 0.007057623006403446, + 0.0694466307759285, + 0.008233325555920601, + 0.011935094371438026, + 0.006981914862990379, + 0.012956881895661354, + 0.07907562702894211, + 0.0082673579454422, + 0.016045695170760155, + 0.008294831961393356, + 0.006792377680540085, + 0.008010968565940857, + 0.006290304474532604, + 0.008651575073599815, + 0.006132049951702356, + 0.005058366805315018, + 0.006544727832078934, + 0.006909546907991171, + 0.006090627983212471, + 0.0041366019286215305, + 0.0059256707318127155, + 0.008588086813688278, + 0.004973345436155796, + 0.005123716313391924, + 0.004842942114919424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.0620332732796669 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629777304828167 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629777304828167, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629777304828167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09481041133403778, + 0.05149480700492859, + 0.08352500200271606, + 0.0433354414999485, + 0.06191675737500191, + 0.035220641642808914, + 0.04918336868286133, + 0.3907654285430908, + 0.06246308982372284, + 0.12728498876094818 + ], + "credit_map": [ + 0.010732702910900116, + 0.008399426005780697, + 0.008691919967532158, + 0.013938972726464272, + 0.01313460897654295, + 0.01279456727206707, + 0.012339147739112377, + 0.014779062941670418, + 0.008071552030742168, + 0.003933387342840433, + 0.010497940704226494, + 0.00583218177780509, + 0.003921086899936199, + 0.007662476040422916, + 0.0048257093876600266, + 0.006750472821295261, + 0.004000070039182901, + 0.0030085656326264143, + 0.00385990715585649, + 0.002970833797007799, + 0.0033148808870464563, + 0.03249774128198624, + 0.0037979772314429283, + 0.030075032263994217, + 0.005511864088475704, + 0.004922560416162014, + 0.006685737520456314, + 0.004395038355141878, + 0.005991171579807997, + 0.0062669808976352215, + 0.00537056103348732, + 0.004191529471427202, + 0.0073927114717662334, + 0.005548419896513224, + 0.0108580831438303, + 0.0048965406604111195, + 0.0067484816536307335, + 0.007907376624643803, + 0.006387215573340654, + 0.012177924625575542, + 0.00449527008458972, + 0.003983906004577875, + 0.006230792962014675, + 0.003267186926677823, + 0.0037593296729028225, + 0.0042806351557374, + 0.0038401775527745485, + 0.00536334328353405, + 0.007799395360052586, + 0.003911800682544708, + 0.012617666274309158, + 0.004980325233191252, + 0.004230160266160965, + 0.005756041966378689, + 0.003915292210876942, + 0.005972689483314753, + 0.14014743268489838, + 0.013675505295395851, + 0.023647310212254524, + 0.012925231829285622, + 0.016123823821544647, + 0.14823439717292786, + 0.008855133317410946, + 0.027156589552760124, + 0.010807067155838013, + 0.007819708436727524, + 0.010196627117693424, + 0.006472288630902767, + 0.007831917144358158, + 0.005848859436810017, + 0.005211362615227699, + 0.008275258354842663, + 0.03514581546187401, + 0.007155317813158035, + 0.01191197894513607, + 0.006926582660526037, + 0.012820150703191757, + 0.011271907016634941, + 0.009216689504683018, + 0.03283655270934105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.007629777304828167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 against a deep, nearly black void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do countless golden points of light appear and converge into a rapidly rotating counterclockwise vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start in a medium shot and slowly advance (zoom in) as the action unfolds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light particles transition from brilliant white to a deep amber while forming a solid fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final form appear as a solid, metallic fist with sharply defined muscles and visible energy lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193772338330746 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193772338330746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193772338330746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03361370787024498, + 0.04101986810564995, + 0.015417599119246006, + 0.02005613036453724, + 0.024311017245054245, + 0.030447913333773613, + 0.06570691615343094, + 0.46638378500938416, + 0.09060896188020706, + 0.2124340832233429 + ], + "credit_map": [ + 0.0044187698513269424, + 0.0036401289980858564, + 0.003365752985700965, + 0.0038741042371839285, + 0.004037945065647364, + 0.0048269107937812805, + 0.00505910674110055, + 0.0043909912928938866, + 0.004725631792098284, + 0.0029098333325237036, + 0.00842207670211792, + 0.0029954854398965836, + 0.003050839761272073, + 0.00680510513484478, + 0.0032819199841469526, + 0.008828973397612572, + 0.0020764973014593124, + 0.0016296406975015998, + 0.0017426119884476066, + 0.0018366437871009111, + 0.0014660650631412864, + 0.0026020468212664127, + 0.002043756889179349, + 0.00202033668756485, + 0.0012680632062256336, + 0.0023852260783314705, + 0.0019394890405237675, + 0.002045502420514822, + 0.0036295426543802023, + 0.0027453775983303785, + 0.0035596159286797047, + 0.002483314834535122, + 0.003343315329402685, + 0.002174953231588006, + 0.0035544822458177805, + 0.001330019673332572, + 0.0024118791334331036, + 0.0036445457953959703, + 0.0023210402578115463, + 0.005530782509595156, + 0.0030022573191672564, + 0.002490296959877014, + 0.0036574562545865774, + 0.0034424415789544582, + 0.003448809962719679, + 0.005342080257833004, + 0.004324967972934246, + 0.0047396039590239525, + 0.009905630722641945, + 0.004813780542463064, + 0.017781050875782967, + 0.006740438751876354, + 0.005027868319302797, + 0.00829196535050869, + 0.005288669373840094, + 0.007857518270611763, + 0.16384954750537872, + 0.015510805882513523, + 0.025586821138858795, + 0.014186380431056023, + 0.01774403266608715, + 0.18513385951519012, + 0.009372561238706112, + 0.034999750554561615, + 0.01486094668507576, + 0.011855295859277248, + 0.0141869792714715, + 0.010198918171226978, + 0.011905240826308727, + 0.009555741213262081, + 0.00741274980828166, + 0.010633096098899841, + 0.026083191856741905, + 0.030059218406677246, + 0.0212269127368927, + 0.03006557561457157, + 0.037893734872341156, + 0.021820519119501114, + 0.02623443305492401, + 0.019050508737564087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.014193772338330746 + } + ], + "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/0000131/group_10/index_tensor([9686, 9686])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_tensor([9686, 9686])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0b3c916332343f4b165c6533ac683b992bb73188 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_tensor([9686, 9686])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from an upward angle, slowly and steadily pans forward, focusing on an ancient soldier perched atop a desolate mountain ridge. Beside him stands a weathered beacon tower, constructed of massive stone blocks, its surface covered in cracks and moss, a silent witness to the howling wind. The soldier's long, black hair is tied back, strands whipping wildly against his weathered face in the biting wind. He wears a suit of worn, dark iron scale armor, its plates scarred by past battles. The hem of his dark, coarse linen tabard flaps violently in the wind. His feet, firmly planted on the rough ground, he draws a massive longbow made of dark, knotty wood with controlled force, the bow creaking mournfully under the immense tension. As the bowstring is drawn to his cheek, the muscles of his arms and back visibly tense beneath his armor, making him look like a statue poised for action. His brow furrowed, his lips pursed, his eyes narrowed in the gusting wind, but his cold, resolute gaze, sharp as an eagle's, fixed its gaze firmly on an invisible target beyond the camera. In the sky, thick, dark clouds rolled rapidly, and the pale, setting sun struggled to cast a faint, cold light, painting the entire scene with a dominant gray-blue hue. The light outlined the cold metal contours of the armor and the rough texture of the rocks, creating an epic atmosphere of loneliness, awe, and resolute will.", + "seed": 32974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_9686_seed_32974_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": -0.030008193850517274, + "vlm_logprob": -0.030008193850517274, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.030008193850517274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015267412178218365 + ], + "opposite_logprob_tokens": [ + -4.890267372131348 + ], + "logprob_reward": -0.015267412178218365, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.015267412178218365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12230879813432693, + 0.03478396683931351, + 0.409976989030838, + 0.05727649852633476, + 0.051705989986658096, + 0.03574846684932709, + 0.038549210876226425, + 0.09745807945728302, + 0.0996910110116005, + 0.05250098928809166 + ], + "credit_map": [ + 0.007796792779117823, + 0.011290247552096844, + 0.02537609077990055, + 0.020910367369651794, + 0.008938340470194817, + 0.01333116926252842, + 0.019842863082885742, + 0.014822924509644508, + 0.002663906430825591, + 0.0028428116347640753, + 0.0040321084670722485, + 0.004905246663838625, + 0.003510858165100217, + 0.005945730954408646, + 0.005877005401998758, + 0.005006302148103714, + 0.13852611184120178, + 0.015319029800593853, + 0.1372898817062378, + 0.02765161544084549, + 0.015677671879529953, + 0.021084142848849297, + 0.03401448577642441, + 0.020414067432284355, + 0.00766033586114645, + 0.004451201297342777, + 0.012663482688367367, + 0.005645486060529947, + 0.0042673200368881226, + 0.008806710131466389, + 0.005695192608982325, + 0.008086769841611385, + 0.010500534437596798, + 0.006453041918575764, + 0.005081041716039181, + 0.004597355145961046, + 0.009875765070319176, + 0.004527846816927195, + 0.006395953241735697, + 0.004274456761777401, + 0.0048546986654400826, + 0.00432591000571847, + 0.002338846679776907, + 0.003982058260589838, + 0.005146292503923178, + 0.006800434086471796, + 0.0051427967846393585, + 0.0031574300955981016, + 0.004402410238981247, + 0.004475078079849482, + 0.006040552165359259, + 0.004142974503338337, + 0.00403685262426734, + 0.004785592667758465, + 0.0072548179887235165, + 0.0034109337721019983, + 0.02496820129454136, + 0.005721620284020901, + 0.007503999397158623, + 0.0056769996881484985, + 0.004257094115018845, + 0.014521564356982708, + 0.010460331104695797, + 0.024348268285393715, + 0.018220312893390656, + 0.008447366766631603, + 0.009458241052925587, + 0.009944921359419823, + 0.00879165343940258, + 0.01669812947511673, + 0.01242110412567854, + 0.01570928283035755, + 0.0073698051273822784, + 0.008150347508490086, + 0.005266243126243353, + 0.006308596581220627, + 0.00694629829376936, + 0.004377733916044235, + 0.007528366521000862, + 0.006553595419973135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.015267412178218365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307778552174568 + ], + "opposite_logprob_tokens": [ + -3.1680777072906494 + ], + "logprob_reward": -0.04307778552174568, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307778552174568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05589621514081955, + 0.02355596236884594, + 0.22481268644332886, + 0.033477552235126495, + 0.04348320513963699, + 0.027561111375689507, + 0.060822539031505585, + 0.03518981859087944, + 0.45610010623931885, + 0.039100758731365204 + ], + "credit_map": [ + 0.006338969338685274, + 0.005916759837418795, + 0.006096609402447939, + 0.0092522157356143, + 0.004870997276157141, + 0.007669062353670597, + 0.009027860127389431, + 0.006723741535097361, + 0.0019924717489629984, + 0.002869208110496402, + 0.003979077097028494, + 0.003143436275422573, + 0.0024178368039429188, + 0.002667085500434041, + 0.003924256656318903, + 0.0025625901762396097, + 0.08704867213964462, + 0.007592091336846352, + 0.08762156218290329, + 0.010146918706595898, + 0.005848728585988283, + 0.007406412158161402, + 0.009695375338196754, + 0.009452922269701958, + 0.004761736840009689, + 0.002959800884127617, + 0.00539135979488492, + 0.003667042125016451, + 0.002747795544564724, + 0.005322284530848265, + 0.0037429891526699066, + 0.004884541500359774, + 0.004453092347830534, + 0.007210223935544491, + 0.002829906065016985, + 0.0034441545139998198, + 0.01188021432608366, + 0.0044876085594296455, + 0.0050254398956894875, + 0.00415256479755044, + 0.00390624045394361, + 0.003649126272648573, + 0.002450311556458473, + 0.0024897195398807526, + 0.004527051467448473, + 0.004858699161559343, + 0.0032064029946923256, + 0.0024735592305660248, + 0.006911490578204393, + 0.004819427151232958, + 0.004213315900415182, + 0.00769426766782999, + 0.007132227066904306, + 0.0054330360144376755, + 0.016992034390568733, + 0.007626735605299473, + 0.0029960921965539455, + 0.0037052177358418703, + 0.0031432437244802713, + 0.0029169050976634026, + 0.002552863908931613, + 0.0052971066907048225, + 0.0045452360063791275, + 0.010033155791461468, + 0.07845693826675415, + 0.045324087142944336, + 0.041627854108810425, + 0.049576498568058014, + 0.04365670680999756, + 0.06445404887199402, + 0.054363492876291275, + 0.07864048331975937, + 0.005406788084656, + 0.005154441110789776, + 0.0036849607713520527, + 0.005540251731872559, + 0.005088352132588625, + 0.0033366114366799593, + 0.005725468043237925, + 0.005163885187357664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.04307778552174568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.035070132464170456, + 0.02662983536720276, + 0.36052700877189636, + 0.028607981279492378, + 0.06751799583435059, + 0.026406949386000633, + 0.08229626715183258, + 0.28715479373931885, + 0.03478274121880531, + 0.051006291061639786 + ], + "credit_map": [ + 0.003005029633641243, + 0.003762808395549655, + 0.006317100487649441, + 0.004905773792415857, + 0.002925611799582839, + 0.0054411813616752625, + 0.003715083934366703, + 0.004997540265321732, + 0.0027580291498452425, + 0.0027437880635261536, + 0.0033964444883167744, + 0.0032865814864635468, + 0.002666936255991459, + 0.003683598944917321, + 0.003931781742721796, + 0.004162674769759178, + 0.13531255722045898, + 0.011579698882997036, + 0.13625161349773407, + 0.017821839079260826, + 0.011116193607449532, + 0.01375800184905529, + 0.01888849213719368, + 0.01579858921468258, + 0.004419388715177774, + 0.002190894912928343, + 0.004281331319361925, + 0.0028360625728964806, + 0.0028107629623264074, + 0.004193669185042381, + 0.004440585616976023, + 0.0034352869261056185, + 0.011798106133937836, + 0.007803814020007849, + 0.00958266295492649, + 0.00841008685529232, + 0.008789489977061749, + 0.005956603214144707, + 0.009894373826682568, + 0.005282859783619642, + 0.0031768223270773888, + 0.0033398128580302, + 0.0025873573031276464, + 0.0034573813900351524, + 0.004573897924274206, + 0.003745686262845993, + 0.003183548105880618, + 0.0023424429818987846, + 0.011395459994673729, + 0.009645141661167145, + 0.004010709002614021, + 0.007593751884996891, + 0.012958143837749958, + 0.019258981570601463, + 0.011080914177000523, + 0.006353164091706276, + 0.08017908036708832, + 0.017747191712260246, + 0.02348734438419342, + 0.01616341434419155, + 0.013734926469624043, + 0.07083529978990555, + 0.029541708528995514, + 0.03546581417322159, + 0.007548448629677296, + 0.0030587029177695513, + 0.0034247960429638624, + 0.003368245903402567, + 0.0028965815436095, + 0.006887107156217098, + 0.0034835587721318007, + 0.004115298856049776, + 0.005819719284772873, + 0.005955438129603863, + 0.005562640726566315, + 0.007074070628732443, + 0.006245983298867941, + 0.0035901374649256468, + 0.008032514713704586, + 0.008725786581635475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.008639458566904068 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094076342880726 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094076342880726, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094076342880726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048036739230155945, + 0.019594866782426834, + 0.48330944776535034, + 0.06422534584999084, + 0.031369712203741074, + 0.07697968184947968, + 0.07433677464723587, + 0.07625061273574829, + 0.042210981249809265, + 0.08368585258722305 + ], + "credit_map": [ + 0.004576690960675478, + 0.003745933761820197, + 0.008564887568354607, + 0.0072357081808149815, + 0.0036787844728678465, + 0.007132895290851593, + 0.006867789663374424, + 0.006234042812138796, + 0.0019529906567186117, + 0.002023833803832531, + 0.002420274540781975, + 0.002611573552712798, + 0.0030207771342247725, + 0.0021087855566293, + 0.0031904326751828194, + 0.0022661988623440266, + 0.18559715151786804, + 0.016288824379444122, + 0.18469718098640442, + 0.022824469953775406, + 0.013102179393172264, + 0.01713208667933941, + 0.023513028398156166, + 0.020154494792222977, + 0.014441248029470444, + 0.004830722697079182, + 0.013079159893095493, + 0.007974568754434586, + 0.004238721914589405, + 0.006251403130590916, + 0.007118533365428448, + 0.006290988530963659, + 0.0036236168816685677, + 0.004185996484011412, + 0.004291378427296877, + 0.0029358386527746916, + 0.005905580706894398, + 0.003671934362500906, + 0.0031316601671278477, + 0.0036237062886357307, + 0.009951815940439701, + 0.009814760647714138, + 0.011104189790785313, + 0.006240701302886009, + 0.009234367869794369, + 0.017825789749622345, + 0.006333209108561277, + 0.006474838126450777, + 0.015620695427060127, + 0.005634499713778496, + 0.003850611625239253, + 0.005947921425104141, + 0.007699102628976107, + 0.01921994425356388, + 0.010033094324171543, + 0.006330896634608507, + 0.019998716190457344, + 0.006459654774516821, + 0.0062604984268546104, + 0.0062487018294632435, + 0.0046523334458470345, + 0.019489644095301628, + 0.0066557759419083595, + 0.006485289894044399, + 0.006167138926684856, + 0.004506262484937906, + 0.005245095584541559, + 0.004290523938834667, + 0.004203047603368759, + 0.003819114062935114, + 0.005834777839481831, + 0.008145018480718136, + 0.011902580969035625, + 0.012636390514671803, + 0.009533249773085117, + 0.015591848641633987, + 0.010047411546111107, + 0.007025239057838917, + 0.009809224866330624, + 0.007139899302273989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.004094076342880726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896223664283752 + ], + "opposite_logprob_tokens": [ + -2.5789623260498047 + ], + "logprob_reward": -0.07896223664283752, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896223664283752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046270355582237244, + 0.01758839562535286, + 0.5900738835334778, + 0.037650350481271744, + 0.018740462139248848, + 0.021827377378940582, + 0.09902343153953552, + 0.07782801240682602, + 0.057821664959192276, + 0.03317609801888466 + ], + "credit_map": [ + 0.004674145951867104, + 0.0035783203784376383, + 0.010213164612650871, + 0.005607425235211849, + 0.003552260110154748, + 0.005475861020386219, + 0.006460990756750107, + 0.006708186585456133, + 0.0020717154256999493, + 0.0017544793663546443, + 0.0029279186856001616, + 0.0018842658028006554, + 0.0017432676395401359, + 0.0022519349586218596, + 0.002342808758839965, + 0.0026120038237422705, + 0.21039332449436188, + 0.020111307501792908, + 0.21091270446777344, + 0.03472686931490898, + 0.0194931048899889, + 0.026265308260917664, + 0.04087045416235924, + 0.027300793677568436, + 0.005332335364073515, + 0.0041693393141031265, + 0.0066754138097167015, + 0.004831693600863218, + 0.0035864778328686953, + 0.004098193719983101, + 0.0035960411187261343, + 0.005360853858292103, + 0.002163416938856244, + 0.0026002004742622375, + 0.0019949637353420258, + 0.001935778884217143, + 0.002969554392620921, + 0.002534332452341914, + 0.002064943080767989, + 0.002477273577824235, + 0.0020393638405948877, + 0.003626076737418771, + 0.002279844833537936, + 0.0027733880560845137, + 0.004344942979514599, + 0.003142566652968526, + 0.0023588829208165407, + 0.0012623114744201303, + 0.017967112362384796, + 0.007371535524725914, + 0.0060808295384049416, + 0.009418737143278122, + 0.008631906472146511, + 0.029210302978754044, + 0.011796155944466591, + 0.008546844124794006, + 0.013892704620957375, + 0.0069733960554003716, + 0.00646998593583703, + 0.008062643930315971, + 0.0046707941219210625, + 0.005755985155701637, + 0.009420896880328655, + 0.022581608965992928, + 0.01525622233748436, + 0.004734347108751535, + 0.004549186676740646, + 0.005956404842436314, + 0.0042876568622887135, + 0.013832076452672482, + 0.005532815586775541, + 0.003672955557703972, + 0.004751307424157858, + 0.005579899065196514, + 0.0034122744109481573, + 0.005133457016199827, + 0.0043306779116392136, + 0.0025227784644812346, + 0.004300524014979601, + 0.0031451797112822533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.07896223664283752 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an upward angle, slowly and steadily pans forward, focusing on an ancient soldier perched atop a desolate mountain ridge. Beside him stands a weathered beacon tower, constructed of massive stone blocks, its surface covered in cracks and moss, a silent witness to the howling wind. The soldier's long, black hair is tied back, strands whipping wildly against his weathered face in the biting wind. He wears a suit of worn, dark iron scale armor, its plates scarred by past battles. The hem of his dark, coarse linen tabard flaps violently in the wind. His feet, firmly planted on the rough ground, he draws a massive longbow made of dark, knotty wood with controlled force, the bow creaking mournfully under the immense tension. As the bowstring is drawn to his cheek, the muscles of his arms and back visibly tense beneath his armor, making him look like a statue poised for action. His brow furrowed, his lips pursed, his eyes narrowed in the gusting wind, but his cold, resolute gaze, sharp as an eagle's, fixed its gaze firmly on an invisible target beyond the camera. In the sky, thick, dark clouds rolled rapidly, and the pale, setting sun struggled to cast a faint, cold light, painting the entire scene with a dominant gray-blue hue. The light outlined the cold metal contours of the armor and the rough texture of the rocks, creating an epic atmosphere of loneliness, awe, and resolute will.", + "seed": 32975, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_9686_seed_32975_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": -0.03796827206388116, + "vlm_logprob": -0.03796827206388116, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03796827206388116, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02118704840540886 + ], + "opposite_logprob_tokens": [ + -4.771186828613281 + ], + "logprob_reward": -0.02118704840540886, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.02118704840540886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2357764095067978, + 0.1162404865026474, + 0.12391907721757889, + 0.05697879567742348, + 0.07003489136695862, + 0.03955231234431267, + 0.04220427945256233, + 0.10333901643753052, + 0.0695655569434166, + 0.1423891931772232 + ], + "credit_map": [ + 0.017340173944830894, + 0.015918778255581856, + 0.05357252433896065, + 0.03224802762269974, + 0.013670328073203564, + 0.03750704601407051, + 0.030217748135328293, + 0.035301774740219116, + 0.011779353953897953, + 0.006785007193684578, + 0.025767283514142036, + 0.010548911988735199, + 0.0069107189774513245, + 0.018521748483181, + 0.014752381481230259, + 0.02117508463561535, + 0.012676050886511803, + 0.004840211942791939, + 0.029911821708083153, + 0.011355913244187832, + 0.005370073951780796, + 0.010573736391961575, + 0.01685228943824768, + 0.032338984310626984, + 0.007733081467449665, + 0.006167836952954531, + 0.008840846829116344, + 0.006692233495414257, + 0.0033673420548439026, + 0.007703382056206465, + 0.0057565295137465, + 0.010717542842030525, + 0.007374269887804985, + 0.00518764415755868, + 0.006683473940938711, + 0.006650405470281839, + 0.004699326120316982, + 0.008896690793335438, + 0.008460046723484993, + 0.022083032876253128, + 0.00509657571092248, + 0.003299109637737274, + 0.005569601897150278, + 0.0035070290323346853, + 0.003515082411468029, + 0.0039040688425302505, + 0.004493414890021086, + 0.010167432948946953, + 0.006203171797096729, + 0.0057344394735991955, + 0.004846974276006222, + 0.0037646740674972534, + 0.009636393748223782, + 0.0035904068499803543, + 0.0035990735050290823, + 0.004829143173992634, + 0.02793212980031967, + 0.009575930424034595, + 0.03676735982298851, + 0.007105544675141573, + 0.005215130280703306, + 0.006896206643432379, + 0.005431384779512882, + 0.004415329545736313, + 0.01034572534263134, + 0.008033380843698978, + 0.010961725376546383, + 0.006696674041450024, + 0.009390800260007381, + 0.009833412244915962, + 0.008229808881878853, + 0.006074029020965099, + 0.013485998846590519, + 0.015107142738997936, + 0.020124070346355438, + 0.02846146933734417, + 0.016002388671040535, + 0.017341244965791702, + 0.01679510623216629, + 0.015071754343807697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.02118704840540886 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496479198336601 + ], + "opposite_logprob_tokens": [ + -2.929964780807495 + ], + "logprob_reward": -0.05496479198336601, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496479198336601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3066161572933197, + 0.08900243788957596, + 0.08788733184337616, + 0.04438313841819763, + 0.07533932477235794, + 0.0465778112411499, + 0.07941536605358124, + 0.12061187624931335, + 0.06930536031723022, + 0.0808611512184143 + ], + "credit_map": [ + 0.037940721958875656, + 0.012571977451443672, + 0.049526866525411606, + 0.03495766222476959, + 0.012921864166855812, + 0.032762035727500916, + 0.03381951153278351, + 0.09211552143096924, + 0.007620126008987427, + 0.00413143914192915, + 0.022988833487033844, + 0.011125265620648861, + 0.004835454281419516, + 0.01198524422943592, + 0.010810792446136475, + 0.015505289658904076, + 0.005874686408787966, + 0.004863047506660223, + 0.013561553321778774, + 0.013459947891533375, + 0.0058535863645374775, + 0.008019831962883472, + 0.019254373386502266, + 0.017000317573547363, + 0.004165830090641975, + 0.004076754674315453, + 0.005286673083901405, + 0.009337115101516247, + 0.0035281835589557886, + 0.005036008544266224, + 0.0064582559280097485, + 0.006494324654340744, + 0.009111936204135418, + 0.008401407860219479, + 0.005870576947927475, + 0.01753983274102211, + 0.006714753806591034, + 0.006575627252459526, + 0.008358226157724857, + 0.012766968458890915, + 0.006793531123548746, + 0.004966788925230503, + 0.007198731880635023, + 0.0037353956140577793, + 0.005520861130207777, + 0.004693278111517429, + 0.004416167736053467, + 0.00925306137651205, + 0.013060398399829865, + 0.008563672192394733, + 0.012331953272223473, + 0.009542996995151043, + 0.01604672707617283, + 0.00698208250105381, + 0.006875493563711643, + 0.006012046243995428, + 0.019286606460809708, + 0.01227980200201273, + 0.01617572084069252, + 0.00959092192351818, + 0.015539904125034809, + 0.024544307962059975, + 0.012258987873792648, + 0.010935644619166851, + 0.014109669253230095, + 0.008561412803828716, + 0.011258387006819248, + 0.006217468064278364, + 0.008509218692779541, + 0.008932674303650856, + 0.006241674534976482, + 0.005474859848618507, + 0.009480471722781658, + 0.009235819801688194, + 0.014726497232913971, + 0.012411343865096569, + 0.00835201982408762, + 0.012404165230691433, + 0.006680701393634081, + 0.007570142857730389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.05496479198336601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.5077998042106628, + 0.05775488168001175, + 0.04904903098940849, + 0.0442873053252697, + 0.050489623099565506, + 0.04187643155455589, + 0.030610468238592148, + 0.14033840596675873, + 0.03155536949634552, + 0.0462387353181839 + ], + "credit_map": [ + 0.07696261256933212, + 0.017168281599879265, + 0.05243009328842163, + 0.03992832824587822, + 0.014171297661960125, + 0.03952362388372421, + 0.06025107204914093, + 0.20736446976661682, + 0.003933575004339218, + 0.00283393869176507, + 0.01858808659017086, + 0.006000679451972246, + 0.002765566809102893, + 0.006713874638080597, + 0.005583215970546007, + 0.011335939168930054, + 0.004195314832031727, + 0.0026997870299965143, + 0.0046190922148525715, + 0.007602022495120764, + 0.00425673695281148, + 0.01034755539149046, + 0.007648078724741936, + 0.0076804435811936855, + 0.00496174581348896, + 0.003832106711342931, + 0.005045856349170208, + 0.004992494825273752, + 0.004965349566191435, + 0.006707340478897095, + 0.005151626653969288, + 0.00863078236579895, + 0.006229709833860397, + 0.0035841846838593483, + 0.005037111230194569, + 0.0048286328092217445, + 0.004527244716882706, + 0.006681504659354687, + 0.005838125944137573, + 0.013763103634119034, + 0.006125933490693569, + 0.006722083780914545, + 0.007023571990430355, + 0.0044354940764606, + 0.005727437790483236, + 0.0025806035846471786, + 0.0037834709510207176, + 0.005477830767631531, + 0.004265058785676956, + 0.0037658936344087124, + 0.0045854272320866585, + 0.0035721624735742807, + 0.005426937714219093, + 0.0021189539693295956, + 0.0031036497093737125, + 0.003772385884076357, + 0.03772569075226784, + 0.014694475568830967, + 0.04747522994875908, + 0.009643401950597763, + 0.009087273851037025, + 0.007558077108114958, + 0.008222115226089954, + 0.005932141561061144, + 0.0036259524058550596, + 0.003733681282028556, + 0.005558942910283804, + 0.003508857451379299, + 0.004102249629795551, + 0.0037227545399218798, + 0.004446978215128183, + 0.002855948405340314, + 0.004897705279290676, + 0.007544622756540775, + 0.007015432231128216, + 0.00978915300220251, + 0.004011760465800762, + 0.00334736006334424, + 0.005208959337323904, + 0.004423736594617367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.00978152547031641 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615036141127348 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003615036141127348, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615036141127348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12702398002147675, + 0.08189338445663452, + 0.11744952946901321, + 0.0382777601480484, + 0.08946412056684494, + 0.045122016221284866, + 0.08256002515554428, + 0.05189444124698639, + 0.13658758997917175, + 0.22972716391086578 + ], + "credit_map": [ + 0.015047872439026833, + 0.009626644663512707, + 0.024234257638454437, + 0.021286599338054657, + 0.00888849887996912, + 0.020807476714253426, + 0.012202211655676365, + 0.014930429868400097, + 0.010219672694802284, + 0.004181531723588705, + 0.025478895753622055, + 0.007026161998510361, + 0.0032004988752305508, + 0.009950145147740841, + 0.007297677453607321, + 0.014538799412548542, + 0.013754075393080711, + 0.005497488658875227, + 0.024365272372961044, + 0.014105024747550488, + 0.004973127972334623, + 0.01793750189244747, + 0.008767039515078068, + 0.028050001710653305, + 0.006225852761417627, + 0.003901597810909152, + 0.008485030382871628, + 0.004316066391766071, + 0.002272180747240782, + 0.004145923536270857, + 0.003555965144187212, + 0.00537514453753829, + 0.01168936025351286, + 0.011798804625868797, + 0.008214607834815979, + 0.018752513453364372, + 0.00560419075191021, + 0.007356331683695316, + 0.009495035745203495, + 0.016553282737731934, + 0.0057990397326648235, + 0.0028779443819075823, + 0.007120891008526087, + 0.0027258831541985273, + 0.0037792951334267855, + 0.004629922565072775, + 0.0053734807297587395, + 0.012815561145544052, + 0.0199753250926733, + 0.009952181950211525, + 0.015719447284936905, + 0.007761718239635229, + 0.005923337768763304, + 0.00683983089402318, + 0.008137858472764492, + 0.008250326849520206, + 0.01078820414841175, + 0.005852939561009407, + 0.012002968229353428, + 0.005960369482636452, + 0.004407915286719799, + 0.004041050095111132, + 0.00497340178117156, + 0.0038675894029438496, + 0.02376975305378437, + 0.019068801775574684, + 0.03383895009756088, + 0.01597817987203598, + 0.011871276423335075, + 0.01072087325155735, + 0.013820645399391651, + 0.0075191170908510685, + 0.0255607720464468, + 0.03128736838698387, + 0.034236930310726166, + 0.049667246639728546, + 0.017823949456214905, + 0.017059821635484695, + 0.028592023998498917, + 0.025499047711491585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.003615036141127348 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029295831918716 + ], + "opposite_logprob_tokens": [ + -2.350292921066284 + ], + "logprob_reward": -0.10029295831918716, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029295831918716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19777584075927734, + 0.0851915180683136, + 0.12615716457366943, + 0.07242397218942642, + 0.06736234575510025, + 0.026725605130195618, + 0.08334451168775558, + 0.18149054050445557, + 0.07377202808856964, + 0.08575651049613953 + ], + "credit_map": [ + 0.024309467524290085, + 0.012843118980526924, + 0.02998085878789425, + 0.026278527453541756, + 0.01116968598216772, + 0.021340373903512955, + 0.02406005561351776, + 0.04779374971985817, + 0.005649127997457981, + 0.007091539911925793, + 0.018522679805755615, + 0.009377806447446346, + 0.004476385191082954, + 0.010460098274052143, + 0.014256944879889488, + 0.015356932766735554, + 0.00952727161347866, + 0.005397704429924488, + 0.027219533920288086, + 0.016561513766646385, + 0.005524585023522377, + 0.010454429313540459, + 0.01700468547642231, + 0.0344674289226532, + 0.007852239534258842, + 0.010051288641989231, + 0.008874726481735706, + 0.014203826896846294, + 0.004395720548927784, + 0.008461986668407917, + 0.007450189907103777, + 0.011133997701108456, + 0.00815737433731556, + 0.010537155903875828, + 0.0062622614204883575, + 0.007912833243608475, + 0.0059465933591127396, + 0.008359488099813461, + 0.006675350945442915, + 0.0135112926363945, + 0.0038395740557461977, + 0.002525866962969303, + 0.005969282239675522, + 0.0029988866299390793, + 0.0023515974171459675, + 0.0018864187877625227, + 0.002508287550881505, + 0.004645690321922302, + 0.016895653679966927, + 0.01011412963271141, + 0.013492142781615257, + 0.0065413424745202065, + 0.01215142011642456, + 0.006318032741546631, + 0.007791847921907902, + 0.010039939545094967, + 0.05357254669070244, + 0.014997614547610283, + 0.07230128347873688, + 0.009229140356183052, + 0.006247540935873985, + 0.010476418770849705, + 0.00843662116676569, + 0.006229385267943144, + 0.009930693544447422, + 0.009260479360818863, + 0.013695542700588703, + 0.008179869502782822, + 0.009029057808220387, + 0.009076055139303207, + 0.008533956483006477, + 0.006066374480724335, + 0.009581750258803368, + 0.009847025386989117, + 0.017674656584858894, + 0.016198113560676575, + 0.008184227161109447, + 0.008255509659647942, + 0.007453065365552902, + 0.008562167175114155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.10029295831918716 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an upward angle, slowly and steadily pans forward, focusing on an ancient soldier perched atop a desolate mountain ridge. Beside him stands a weathered beacon tower, constructed of massive stone blocks, its surface covered in cracks and moss, a silent witness to the howling wind. The soldier's long, black hair is tied back, strands whipping wildly against his weathered face in the biting wind. He wears a suit of worn, dark iron scale armor, its plates scarred by past battles. The hem of his dark, coarse linen tabard flaps violently in the wind. His feet, firmly planted on the rough ground, he draws a massive longbow made of dark, knotty wood with controlled force, the bow creaking mournfully under the immense tension. As the bowstring is drawn to his cheek, the muscles of his arms and back visibly tense beneath his armor, making him look like a statue poised for action. His brow furrowed, his lips pursed, his eyes narrowed in the gusting wind, but his cold, resolute gaze, sharp as an eagle's, fixed its gaze firmly on an invisible target beyond the camera. In the sky, thick, dark clouds rolled rapidly, and the pale, setting sun struggled to cast a faint, cold light, painting the entire scene with a dominant gray-blue hue. The light outlined the cold metal contours of the armor and the rough texture of the rocks, creating an epic atmosphere of loneliness, awe, and resolute will.", + "seed": 32976, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_9686_seed_32976_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": -0.03098709722980857, + "vlm_logprob": -0.03098709722980857, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03098709722980857, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02398231439292431 + ], + "opposite_logprob_tokens": [ + -4.273982524871826 + ], + "logprob_reward": -0.02398231439292431, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02398231439292431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1287418007850647, + 0.10075673460960388, + 0.08159887790679932, + 0.09062904864549637, + 0.049562521278858185, + 0.3270353078842163, + 0.04038098827004433, + 0.03159637749195099, + 0.06102576479315758, + 0.08867250382900238 + ], + "credit_map": [ + 0.0175545085221529, + 0.012783926911652088, + 0.018830977380275726, + 0.010187496431171894, + 0.013672924600541592, + 0.02070445939898491, + 0.011579143814742565, + 0.023428374901413918, + 0.016767680644989014, + 0.010500597767531872, + 0.009968745522201061, + 0.004220828413963318, + 0.012418323196470737, + 0.03373131528496742, + 0.004814742598682642, + 0.008334511891007423, + 0.031119296327233315, + 0.008223060518503189, + 0.00830097496509552, + 0.005674748215824366, + 0.008028870448470116, + 0.012671045958995819, + 0.002909245667979121, + 0.004671634640544653, + 0.014224986545741558, + 0.01484751794487238, + 0.006351367570459843, + 0.0042069293558597565, + 0.012702271342277527, + 0.02916312776505947, + 0.004388389643281698, + 0.004744452890008688, + 0.0056457738392055035, + 0.007447308395057917, + 0.00653168885037303, + 0.00609247712418437, + 0.006551917176693678, + 0.005557663273066282, + 0.006988190580159426, + 0.0047475057654082775, + 0.06938615441322327, + 0.010859583504498005, + 0.13018588721752167, + 0.017087532207369804, + 0.011124592274427414, + 0.01710423454642296, + 0.043856505304574966, + 0.027430854737758636, + 0.005322488956153393, + 0.003201302606612444, + 0.010365819558501244, + 0.0037341462448239326, + 0.0032873679883778095, + 0.005747827235609293, + 0.0049620820209383965, + 0.003759954124689102, + 0.0033685388043522835, + 0.00408132141456008, + 0.00469102943316102, + 0.003679710440337658, + 0.0027114320546388626, + 0.004125102423131466, + 0.004413523245602846, + 0.0045257206074893475, + 0.004649467766284943, + 0.009242793545126915, + 0.005483182612806559, + 0.006611164193600416, + 0.006883443333208561, + 0.007513800170272589, + 0.011896822601556778, + 0.008745087310671806, + 0.007677232846617699, + 0.008360876701772213, + 0.0072434223257005215, + 0.00643361359834671, + 0.019802970811724663, + 0.01239758264273405, + 0.01585976779460907, + 0.010897032916545868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02398231439292431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003770768642426 + ], + "opposite_logprob_tokens": [ + -2.695037603378296 + ], + "logprob_reward": -0.07003770768642426, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003770768642426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1248043030500412, + 0.11903117597103119, + 0.10461604595184326, + 0.07656392455101013, + 0.17892566323280334, + 0.08685870468616486, + 0.11170351505279541, + 0.040661655366420746, + 0.07812117040157318, + 0.07871382683515549 + ], + "credit_map": [ + 0.015827564522624016, + 0.01576286181807518, + 0.014672129414975643, + 0.010094273835420609, + 0.021563831716775894, + 0.020517678931355476, + 0.010092117823660374, + 0.01627383381128311, + 0.019337017089128494, + 0.014963546767830849, + 0.008920309133827686, + 0.006195514462888241, + 0.019621890038251877, + 0.03483935818076134, + 0.006385829299688339, + 0.008767712861299515, + 0.03179232403635979, + 0.011356792412698269, + 0.01157635822892189, + 0.008270532824099064, + 0.014811336994171143, + 0.014402724802494049, + 0.005290023982524872, + 0.007115952204912901, + 0.010749248787760735, + 0.013611920177936554, + 0.004892132245004177, + 0.005672836676239967, + 0.010904158465564251, + 0.021630778908729553, + 0.005309780593961477, + 0.0037930654361844063, + 0.018231095746159554, + 0.022047827020287514, + 0.023772643879055977, + 0.022474315017461777, + 0.023451197892427444, + 0.020913271233439445, + 0.03106066584587097, + 0.016974646598100662, + 0.009592387825250626, + 0.010379163548350334, + 0.023411499336361885, + 0.010547488927841187, + 0.006791828665882349, + 0.006666306406259537, + 0.00992146972566843, + 0.009548570029437542, + 0.023977532982826233, + 0.008420063182711601, + 0.027530554682016373, + 0.011855380609631538, + 0.007181236520409584, + 0.01141563430428505, + 0.011131850071251392, + 0.01019127108156681, + 0.0038825341034680605, + 0.0063943020068109035, + 0.006066642235964537, + 0.0048345522955060005, + 0.003599737770855427, + 0.004804241470992565, + 0.004820344969630241, + 0.006259299349039793, + 0.005795447155833244, + 0.012714892625808716, + 0.00621040677651763, + 0.008239728398621082, + 0.008850522339344025, + 0.00905389990657568, + 0.01642179861664772, + 0.010834479704499245, + 0.006431822199374437, + 0.00773930037394166, + 0.007396959234029055, + 0.00861210934817791, + 0.01599629409611225, + 0.010139400139451027, + 0.014005713164806366, + 0.008392224088311195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.07003770768642426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06025930866599083, + 0.046517498791217804, + 0.026908211410045624, + 0.042957913130521774, + 0.02888987399637699, + 0.41685226559638977, + 0.09320265054702759, + 0.17277687788009644, + 0.05852224677801132, + 0.05311320722103119 + ], + "credit_map": [ + 0.006189828272908926, + 0.009880254976451397, + 0.008563407696783543, + 0.007379651535302401, + 0.010288978926837444, + 0.00475964043289423, + 0.005518442951142788, + 0.007679101079702377, + 0.0072657023556530476, + 0.005634598433971405, + 0.004018543753772974, + 0.0029385967645794153, + 0.006561063230037689, + 0.010265033692121506, + 0.0037997120525687933, + 0.006034248974174261, + 0.003973335027694702, + 0.003507064189761877, + 0.002653771545737982, + 0.00247760652564466, + 0.004561251029372215, + 0.003802312072366476, + 0.00286979041993618, + 0.003063078736886382, + 0.008006876334547997, + 0.0044691492803394794, + 0.004269307479262352, + 0.003734280588105321, + 0.005231161136180162, + 0.009078925475478172, + 0.0036569966468960047, + 0.004511210601776838, + 0.003838571486994624, + 0.003419050946831703, + 0.003703361377120018, + 0.0038287839852273464, + 0.0028486293740570545, + 0.004568139091134071, + 0.003153412137180567, + 0.003529922105371952, + 0.09488728642463684, + 0.012367657385766506, + 0.17845883965492249, + 0.017366360872983932, + 0.012980787083506584, + 0.012869448401033878, + 0.0587896890938282, + 0.029132183641195297, + 0.008317035622894764, + 0.005696532782167196, + 0.02362392097711563, + 0.009183661080896854, + 0.007006807252764702, + 0.012546195648610592, + 0.018666930496692657, + 0.00816156156361103, + 0.016792327165603638, + 0.021986454725265503, + 0.027540262788534164, + 0.021617911756038666, + 0.013941345736384392, + 0.021127117797732353, + 0.02223013900220394, + 0.027541300281882286, + 0.005435796454548836, + 0.0068433647975325584, + 0.006375370547175407, + 0.007041387725621462, + 0.0062958006747066975, + 0.00762051111087203, + 0.012173415161669254, + 0.006736595183610916, + 0.0040098219178617, + 0.004408574663102627, + 0.004703106824308634, + 0.008239022456109524, + 0.007161313202232122, + 0.008655558340251446, + 0.00920155830681324, + 0.006734246388077736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "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 scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046386257745325565 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.0046386257745325565, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046386257745325565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0654413178563118, + 0.29574838280677795, + 0.11669206619262695, + 0.09730330109596252, + 0.14161017537117004, + 0.03481196239590645, + 0.0556170828640461, + 0.051422953605651855, + 0.0353810079395771, + 0.10597174614667892 + ], + "credit_map": [ + 0.009447403252124786, + 0.006630134303122759, + 0.01016658078879118, + 0.00494417455047369, + 0.007550302892923355, + 0.010188661515712738, + 0.006799252238124609, + 0.009714804589748383, + 0.04816790670156479, + 0.031215572729706764, + 0.023765958845615387, + 0.017155714333057404, + 0.04133691266179085, + 0.08367270231246948, + 0.017071381211280823, + 0.03336222469806671, + 0.049894195050001144, + 0.010601844638586044, + 0.012826482765376568, + 0.0053874775767326355, + 0.007968676276504993, + 0.01986108347773552, + 0.004394866060465574, + 0.005757431965321302, + 0.016001194715499878, + 0.01308941375464201, + 0.01764051802456379, + 0.0061326404102146626, + 0.012054361402988434, + 0.018978025764226913, + 0.005722419358789921, + 0.007684721611440182, + 0.025753198191523552, + 0.017019249498844147, + 0.018996698781847954, + 0.018693316727876663, + 0.016860559582710266, + 0.018964963033795357, + 0.011651151813566685, + 0.013671036809682846, + 0.0036175623536109924, + 0.003023287979885936, + 0.004543558694422245, + 0.005632392596453428, + 0.0029742259066551924, + 0.00484070461243391, + 0.003926374949514866, + 0.006253852043300867, + 0.011472087353467941, + 0.002946414751932025, + 0.01724035106599331, + 0.005174137651920319, + 0.0027979544829577208, + 0.004906149581074715, + 0.00606142170727253, + 0.005018566735088825, + 0.0046609328128397465, + 0.007722073700278997, + 0.004328970331698656, + 0.005911486689001322, + 0.004773857071995735, + 0.007287136744707823, + 0.007388225756585598, + 0.009350265376269817, + 0.0030955979600548744, + 0.006845845840871334, + 0.0032674353569746017, + 0.004946445114910603, + 0.004110071342438459, + 0.0037765491288155317, + 0.005645858123898506, + 0.003693200182169676, + 0.009427502751350403, + 0.010718288831412792, + 0.009936985559761524, + 0.011198291555047035, + 0.01616307906806469, + 0.016183454543352127, + 0.014741642400622368, + 0.017602495849132538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.0046386257745325565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864410310983658 + ], + "opposite_logprob_tokens": [ + -3.0486440658569336 + ], + "logprob_reward": -0.04864410310983658, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864410310983658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07311012595891953, + 0.058936502784490585, + 0.051882848143577576, + 0.27146708965301514, + 0.03594110906124115, + 0.36091887950897217, + 0.05434470996260643, + 0.021435940638184547, + 0.01332803349941969, + 0.0586346797645092 + ], + "credit_map": [ + 0.006684201769530773, + 0.015074200928211212, + 0.007626370992511511, + 0.008924838155508041, + 0.013107114471495152, + 0.007020521908998489, + 0.007165317889302969, + 0.0075075747445225716, + 0.009700111113488674, + 0.007137890439480543, + 0.005568748340010643, + 0.005096298176795244, + 0.008141465485095978, + 0.012727716006338596, + 0.003920881077647209, + 0.006643397733569145, + 0.009080242365598679, + 0.0041292281821370125, + 0.005561856087297201, + 0.0041863517835736275, + 0.007913317531347275, + 0.01308547705411911, + 0.003678328590467572, + 0.004248055629432201, + 0.04257534444332123, + 0.04312768206000328, + 0.018387164920568466, + 0.013926384039223194, + 0.036447834223508835, + 0.0810123160481453, + 0.019170086830854416, + 0.01682031899690628, + 0.004477375652641058, + 0.0046899099834263325, + 0.004794218577444553, + 0.00430546049028635, + 0.005178259685635567, + 0.00418730778619647, + 0.0034622508101165295, + 0.004846332594752312, + 0.0752117931842804, + 0.01432497613132, + 0.15060506761074066, + 0.017003031447529793, + 0.012869756668806076, + 0.011249496601521969, + 0.05240229144692421, + 0.027252521365880966, + 0.008005562238395214, + 0.0037057106383144855, + 0.016555605456233025, + 0.005491903051733971, + 0.0036419846583157778, + 0.004024459049105644, + 0.007721079979091883, + 0.005198411177843809, + 0.0021297279745340347, + 0.003474160563200712, + 0.0025690116453915834, + 0.00270658265799284, + 0.0018810795154422522, + 0.002628048649057746, + 0.0027257422916591167, + 0.003321591531857848, + 0.0011083504650741816, + 0.0020398187916725874, + 0.001469314331188798, + 0.0017631712835282087, + 0.00143600651063025, + 0.002252645790576935, + 0.0017536274390295148, + 0.0015051019145175815, + 0.004274018574506044, + 0.004334914032369852, + 0.005305348429828882, + 0.004359041340649128, + 0.01037923526018858, + 0.012752803042531013, + 0.010264464654028416, + 0.006964860484004021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.04864410310983658 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from an upward angle, slowly and steadily pans forward, focusing on an ancient soldier perched atop a desolate mountain ridge. Beside him stands a weathered beacon tower, constructed of massive stone blocks, its surface covered in cracks and moss, a silent witness to the howling wind. The soldier's long, black hair is tied back, strands whipping wildly against his weathered face in the biting wind. He wears a suit of worn, dark iron scale armor, its plates scarred by past battles. The hem of his dark, coarse linen tabard flaps violently in the wind. His feet, firmly planted on the rough ground, he draws a massive longbow made of dark, knotty wood with controlled force, the bow creaking mournfully under the immense tension. As the bowstring is drawn to his cheek, the muscles of his arms and back visibly tense beneath his armor, making him look like a statue poised for action. His brow furrowed, his lips pursed, his eyes narrowed in the gusting wind, but his cold, resolute gaze, sharp as an eagle's, fixed its gaze firmly on an invisible target beyond the camera. In the sky, thick, dark clouds rolled rapidly, and the pale, setting sun struggled to cast a faint, cold light, painting the entire scene with a dominant gray-blue hue. The light outlined the cold metal contours of the armor and the rough texture of the rocks, creating an epic atmosphere of loneliness, awe, and resolute will.", + "seed": 32977, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_10/index_9686_seed_32977_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": -0.026924182707443834, + "vlm_logprob": -0.026924182707443834, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.026924182707443834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02018248662352562 + ], + "opposite_logprob_tokens": [ + -4.1451826095581055 + ], + "logprob_reward": -0.02018248662352562, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02018248662352562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1543400138616562, + 0.020370738580822945, + 0.05191901698708534, + 0.037907011806964874, + 0.3813840448856354, + 0.06037672609090805, + 0.15761786699295044, + 0.0540594607591629, + 0.04336462542414665, + 0.03866055607795715 + ], + "credit_map": [ + 0.014440336264669895, + 0.02842271327972412, + 0.01568959653377533, + 0.027864918112754822, + 0.020999720320105553, + 0.019050458446145058, + 0.013055677525699139, + 0.014816590584814548, + 0.0025687136221677065, + 0.0016888772370293736, + 0.0019047600217163563, + 0.0019438593881204724, + 0.0023361349012702703, + 0.0028801076114177704, + 0.0015332754701375961, + 0.00551501102745533, + 0.004880306776612997, + 0.008973944932222366, + 0.0039066593162715435, + 0.007704142481088638, + 0.007858200930058956, + 0.005792365875095129, + 0.009303647093474865, + 0.00349974911659956, + 0.006938816048204899, + 0.002675203839316964, + 0.007086428813636303, + 0.0037308617029339075, + 0.0031709051690995693, + 0.0074348533526062965, + 0.0023559904657304287, + 0.004513952415436506, + 0.16293050348758698, + 0.016574295237660408, + 0.0722542479634285, + 0.015559757128357887, + 0.022064849734306335, + 0.013534492813050747, + 0.07029546797275543, + 0.008170433342456818, + 0.0089633259922266, + 0.005830344744026661, + 0.009529727511107922, + 0.009029436856508255, + 0.0040047429502010345, + 0.005268540233373642, + 0.010817816480994225, + 0.006932794116437435, + 0.015413949266076088, + 0.013001631014049053, + 0.0354352630674839, + 0.023980548605322838, + 0.01065695658326149, + 0.014964749105274677, + 0.023039791733026505, + 0.021124985069036484, + 0.007954763248562813, + 0.0058647748082876205, + 0.003808016423135996, + 0.009949890896677971, + 0.005156118422746658, + 0.004086795728653669, + 0.009657926857471466, + 0.0075811706483364105, + 0.005320128984749317, + 0.004945256281644106, + 0.0032286117784678936, + 0.003815780393779278, + 0.01091236062347889, + 0.005925371777266264, + 0.004969482310116291, + 0.004247630946338177, + 0.004610604606568813, + 0.006173150148242712, + 0.00505042914301157, + 0.00684009725227952, + 0.0030908307526260614, + 0.0028631563764065504, + 0.004710246808826923, + 0.005322037730365992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.02018248662352562 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866942763328552 + ], + "opposite_logprob_tokens": [ + -3.0486693382263184 + ], + "logprob_reward": -0.04866942763328552, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866942763328552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06722128391265869, + 0.026406625285744667, + 0.05087285488843918, + 0.07645174115896225, + 0.3194865584373474, + 0.08305583894252777, + 0.0509926900267601, + 0.12088324874639511, + 0.14178861677646637, + 0.06284058839082718 + ], + "credit_map": [ + 0.011757983826100826, + 0.007552654482424259, + 0.007699229288846254, + 0.012079850770533085, + 0.008307590149343014, + 0.008264502510428429, + 0.005114155821502209, + 0.00644531287252903, + 0.004788833670318127, + 0.002251826459541917, + 0.0025037750601768494, + 0.0033882453572005033, + 0.0029590828344225883, + 0.003297739662230015, + 0.003327201586216688, + 0.003889919025823474, + 0.010083489120006561, + 0.0036564383190125227, + 0.004293693695217371, + 0.005490448791533709, + 0.006087580230087042, + 0.008208057843148708, + 0.004219753667712212, + 0.008833390660583973, + 0.0110237468034029, + 0.010119604878127575, + 0.007815768010914326, + 0.012222114019095898, + 0.012567336671054363, + 0.00734334159642458, + 0.00575020257383585, + 0.009609624743461609, + 0.13668625056743622, + 0.013387800194323063, + 0.057985663414001465, + 0.01268086675554514, + 0.02295151725411415, + 0.011007271707057953, + 0.058206718415021896, + 0.006580468267202377, + 0.009691492654383183, + 0.010691597126424313, + 0.011799433268606663, + 0.013360132463276386, + 0.0057951840572059155, + 0.007639643736183643, + 0.014385335147380829, + 0.009693020023405552, + 0.004394379444420338, + 0.006406261585652828, + 0.011256108060479164, + 0.00861247070133686, + 0.0034880791790783405, + 0.003832446876913309, + 0.006741069257259369, + 0.006261870265007019, + 0.0075385537929832935, + 0.011991189792752266, + 0.01111496239900589, + 0.030263807624578476, + 0.00910297129303217, + 0.007048718631267548, + 0.026199128478765488, + 0.017623912543058395, + 0.01461076084524393, + 0.0148012051358819, + 0.008029061369597912, + 0.017558393999934196, + 0.03812789544463158, + 0.02045964077115059, + 0.015638278797268867, + 0.01256337109953165, + 0.007688353769481182, + 0.007355289533734322, + 0.0072696455754339695, + 0.009950857609510422, + 0.006762489676475525, + 0.005610223859548569, + 0.009309213608503342, + 0.008894513361155987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.04866942763328552 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076326169073581696 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.0076326169073581696, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076326169073581696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0435895211994648, + 0.02141728065907955, + 0.016511106863617897, + 0.021635431796312332, + 0.33656612038612366, + 0.07909715175628662, + 0.11562273651361465, + 0.04776165634393692, + 0.23259179294109344, + 0.08520713448524475 + ], + "credit_map": [ + 0.004048094153404236, + 0.005135504994541407, + 0.004587937146425247, + 0.009253568015992641, + 0.005002951715141535, + 0.0055262427777051926, + 0.004374890588223934, + 0.005660331342369318, + 0.0017494324129074812, + 0.002139483345672488, + 0.0021114484407007694, + 0.003352175699546933, + 0.002687912667170167, + 0.00302309961989522, + 0.002561117522418499, + 0.003792610950767994, + 0.0019237515516579151, + 0.001382456161081791, + 0.0019476276356726885, + 0.0023059675004333258, + 0.0013683534925803542, + 0.002935634693130851, + 0.002294522477313876, + 0.0023527939338237047, + 0.0032524277921766043, + 0.0019260075641795993, + 0.003082155715674162, + 0.0032323782797902822, + 0.0021045380271971226, + 0.003964839968830347, + 0.001427943236194551, + 0.0026451414451003075, + 0.1431208699941635, + 0.01310680154711008, + 0.05758349224925041, + 0.012178857810795307, + 0.030795976519584656, + 0.010660829953849316, + 0.060120292007923126, + 0.008999014273285866, + 0.009694899432361126, + 0.010408601723611355, + 0.005943668074905872, + 0.012824712321162224, + 0.006757067982107401, + 0.009959526360034943, + 0.010686999186873436, + 0.012821678072214127, + 0.011115129105746746, + 0.008906283415853977, + 0.030527008697390556, + 0.015298929996788502, + 0.006568809039890766, + 0.00936079304665327, + 0.019793078303337097, + 0.014052707701921463, + 0.005250232294201851, + 0.005948556121438742, + 0.0026072708424180746, + 0.010138493962585926, + 0.004617972299456596, + 0.004357445053756237, + 0.008309142664074898, + 0.006532543804496527, + 0.022492313757538795, + 0.023378359153866768, + 0.015986470505595207, + 0.026738740503787994, + 0.061450690031051636, + 0.033858589828014374, + 0.02294093556702137, + 0.025745708495378494, + 0.011731206439435482, + 0.011235803365707397, + 0.007212685886770487, + 0.009849883615970612, + 0.01554684154689312, + 0.01051176805049181, + 0.010586276650428772, + 0.008532674051821232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.0076326169073581696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042056355625391006, + 0.035875603556632996, + 0.08967754244804382, + 0.08452821522951126, + 0.3463861644268036, + 0.041745420545339584, + 0.18557323515415192, + 0.0864272490143776, + 0.06740588694810867, + 0.020324312150478363 + ], + "credit_map": [ + 0.005396406166255474, + 0.0035822573117911816, + 0.00487331859767437, + 0.007848968729376793, + 0.004089789465069771, + 0.005089804995805025, + 0.00448929937556386, + 0.006686510518193245, + 0.0067018731497228146, + 0.0026317674200981855, + 0.006399939302355051, + 0.004126681014895439, + 0.0027953891549259424, + 0.005886216182261705, + 0.0021979750599712133, + 0.0051357620395720005, + 0.01847536861896515, + 0.006553830113261938, + 0.008747153915464878, + 0.008390442468225956, + 0.010260336101055145, + 0.015004811808466911, + 0.007534305099397898, + 0.01471130270510912, + 0.01892898604273796, + 0.007459933403879404, + 0.016620010137557983, + 0.005889520980417728, + 0.006754016969352961, + 0.011789768934249878, + 0.004739449825137854, + 0.012346530333161354, + 0.1469205766916275, + 0.014951225370168686, + 0.07240903377532959, + 0.015685059130191803, + 0.008709940128028393, + 0.013716578483581543, + 0.06576227396726608, + 0.00823147315531969, + 0.006622465327382088, + 0.0035913181491196156, + 0.007744932547211647, + 0.0056782374158501625, + 0.0027471352368593216, + 0.002914865966886282, + 0.008385743014514446, + 0.00406072149053216, + 0.020346229895949364, + 0.013412331230938435, + 0.05149519443511963, + 0.025774650275707245, + 0.01053293514996767, + 0.016618335619568825, + 0.025655770674347878, + 0.021737786009907722, + 0.01604677364230156, + 0.0107054952532053, + 0.005776507314294577, + 0.012255236506462097, + 0.00746289175003767, + 0.009473275393247604, + 0.012006490491330624, + 0.012700577266514301, + 0.006647022441029549, + 0.006996539421379566, + 0.004026171751320362, + 0.0069954548962414265, + 0.01817331649363041, + 0.01001899316906929, + 0.007483392488211393, + 0.0070649925619363785, + 0.002828591736033559, + 0.0028551295399665833, + 0.002182543044909835, + 0.0026402613148093224, + 0.0035013887099921703, + 0.0022495253942906857, + 0.0020604440942406654, + 0.002006426453590393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a desolate mountain ridge with a weathered beacon tower made of massive stone blocks nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing dark iron scale armor and a dark coarse linen tabard?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the soldier drawing a massive longbow made of dark knotty wood with the string pulled to his cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with thick dark clouds and a pale setting sun casting a dominant gray-blue hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from an upward angle and slowly panning forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494357645511627 + ], + "opposite_logprob_tokens": [ + -2.929943561553955 + ], + "logprob_reward": -0.05494357645511627, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494357645511627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0975320041179657, + 0.04401620104908943, + 0.08748434484004974, + 0.05376134067773819, + 0.12107321619987488, + 0.11215938627719879, + 0.19419831037521362, + 0.188073068857193, + 0.08293592184782028, + 0.018766187131404877 + ], + "credit_map": [ + 0.009018064476549625, + 0.017075706273317337, + 0.010097024030983448, + 0.01737172342836857, + 0.011758377775549889, + 0.0110087301582098, + 0.008368343114852905, + 0.01283402368426323, + 0.004279197193682194, + 0.006883278023451567, + 0.0032445730175822973, + 0.005866847466677427, + 0.007768343668431044, + 0.005521622020751238, + 0.006019281223416328, + 0.004433053079992533, + 0.011922516860067844, + 0.007317418232560158, + 0.011778630316257477, + 0.01002046000212431, + 0.007778713013976812, + 0.016101185232400894, + 0.009692993015050888, + 0.012872423976659775, + 0.009428412653505802, + 0.005842986050993204, + 0.004881891887634993, + 0.00771155534312129, + 0.008066732436418533, + 0.006502446718513966, + 0.0037456483114510775, + 0.007581659127026796, + 0.03339236602187157, + 0.008234154433012009, + 0.025329047814011574, + 0.006051418371498585, + 0.01742640882730484, + 0.007233935408294201, + 0.01860927976667881, + 0.004796595778316259, + 0.017532167956233025, + 0.013523153960704803, + 0.01443703006953001, + 0.013218021020293236, + 0.009748218581080437, + 0.009431152604520321, + 0.018341660499572754, + 0.015927966684103012, + 0.017367985099554062, + 0.01587461680173874, + 0.04703006148338318, + 0.030833568423986435, + 0.011513073928654194, + 0.01636941358447075, + 0.03216253221035004, + 0.023047061637043953, + 0.021532271057367325, + 0.026153389364480972, + 0.010655826888978481, + 0.03255053609609604, + 0.018409082666039467, + 0.01455098669975996, + 0.03858035057783127, + 0.02564062364399433, + 0.008767243474721909, + 0.009004496037960052, + 0.004869386088103056, + 0.007790177129209042, + 0.02312225103378296, + 0.012765739113092422, + 0.008319132961332798, + 0.008297485299408436, + 0.003310560015961528, + 0.003686932846903801, + 0.0015526197385042906, + 0.003347477177157998, + 0.0023640054278075695, + 0.0017307017697021365, + 0.0014475052012130618, + 0.0013263822766020894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.05494357645511627 + } + ], + "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/0000131/group_11/index_tensor([6739, 6739])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_11/index_tensor([6739, 6739])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b42e4bc55b891d3312f4454c5ff6dbec05c5a476 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_11/index_tensor([6739, 6739])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic overhead shot, beginning with a close-up of the eagle and smoothly zooming out as it moves, captures a majestic eagle perched atop the desolate, windswept summit of a towering mountain. Its snow-white head feathers, piercing, focused golden eyes, and a massive, hooked yellow beak gleam with a cool sheen in the first golden rays of morning sunlight. Its deep brown body is layered with feathers, and its powerful yellow talons grip a ridged, granite rock, shrouded in lingering snow. Behind it, a vast panorama of jagged, snow-capped peaks stretches to the horizon, while wisps of morning mist curl through the valley. After a moment of stillness, the eagle's body suddenly tensed, and then with an explosive yet extremely calm movement, it spread its wide wings steadily and powerfully. The camera then zoomed out to fully display its amazing wingspan. Each main flight feather was outlined with a golden edge in the backlight, as if declaring its sovereignty over the sky. The whole picture was filled with a tranquil atmosphere of majesty, freedom and untamed natural power.", + "seed": 4789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_11/index_6739_seed_4789_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": -0.29484697096049783, + "vlm_logprob": -0.29484697096049783, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.29484697096049783, + "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 on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.2247824668884277 + ], + "opposite_logprob_tokens": [ + -0.34978246688842773 + ], + "logprob_reward": -1.2247824668884277, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2247824668884277, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2721482515335083, + 0.3132780194282532, + 0.0457821786403656, + 0.0618307888507843, + 0.17156267166137695, + 0.02249165251851082, + 0.016022149473428726, + 0.01146235503256321, + 0.02119210548698902, + 0.06422983109951019 + ], + "credit_map": [ + 0.06311489641666412, + 0.03717798367142677, + 0.018939197063446045, + 0.019627798348665237, + 0.035342682152986526, + 0.03375804051756859, + 0.018306579440832138, + 0.045881062746047974, + 0.06487400084733963, + 0.03598567470908165, + 0.04247841611504555, + 0.021442562341690063, + 0.025809647515416145, + 0.059063393622636795, + 0.020053856074810028, + 0.04357043653726578, + 0.010232659056782722, + 0.004815182648599148, + 0.006373139563947916, + 0.002974419854581356, + 0.004461423493921757, + 0.005678670946508646, + 0.0038693479727953672, + 0.007377331610769033, + 0.012496771290898323, + 0.008020981214940548, + 0.005603807978332043, + 0.00402213167399168, + 0.0043867421336472034, + 0.009024390950798988, + 0.0044573405757546425, + 0.013818630017340183, + 0.017847899347543716, + 0.007044103462249041, + 0.0088429544121027, + 0.005550974979996681, + 0.00646247947588563, + 0.06903668493032455, + 0.008241266943514347, + 0.04853631928563118, + 0.005719579290598631, + 0.0024880312848836184, + 0.0027183243073523045, + 0.0017227004282176495, + 0.0024647016543895006, + 0.002584467874839902, + 0.00225629098713398, + 0.0025375557597726583, + 0.0014458831865340471, + 0.0022733050864189863, + 0.0016738434787839651, + 0.0018171039409935474, + 0.0023840414360165596, + 0.0020484921988099813, + 0.0021677997428923845, + 0.002211680868640542, + 0.0012858962872996926, + 0.001731871161609888, + 0.0016362527385354042, + 0.0008529335609637201, + 0.002650005742907524, + 0.0009368581231683493, + 0.0008155823452398181, + 0.0015529552474617958, + 0.002873062388971448, + 0.0034257995430380106, + 0.003422813257202506, + 0.0023197561968117952, + 0.0021744261030107737, + 0.0022306416649371386, + 0.0017714177956804633, + 0.002974187023937702, + 0.008779901079833508, + 0.006971891038119793, + 0.008404739201068878, + 0.006444940809160471, + 0.005992310121655464, + 0.01333136297762394, + 0.006350834853947163, + 0.007953854277729988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -1.2247824668884277 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374454751610756 + ], + "opposite_logprob_tokens": [ + -3.4087445735931396 + ], + "logprob_reward": -0.03374454751610756, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374454751610756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2686219811439514, + 0.14276206493377686, + 0.08518083393573761, + 0.14447592198848724, + 0.1685604751110077, + 0.06946055591106415, + 0.0378570519387722, + 0.016491780057549477, + 0.01429027784615755, + 0.052299097180366516 + ], + "credit_map": [ + 0.05811261758208275, + 0.03850562870502472, + 0.025060612708330154, + 0.01810385100543499, + 0.03247841075062752, + 0.032027289271354675, + 0.017235318198800087, + 0.04709824547171593, + 0.02038397081196308, + 0.015581372193992138, + 0.027514666318893433, + 0.01097910851240158, + 0.015390178188681602, + 0.016510093584656715, + 0.01180926151573658, + 0.024593409150838852, + 0.017795389518141747, + 0.007849938236176968, + 0.012452229857444763, + 0.005772050935775042, + 0.010074606165289879, + 0.011457386426627636, + 0.007551375310868025, + 0.012227851897478104, + 0.027275780215859413, + 0.015212969854474068, + 0.01441000122576952, + 0.008669096976518631, + 0.014361660927534103, + 0.021519381552934647, + 0.010520034469664097, + 0.032506998628377914, + 0.0166485458612442, + 0.0058225407265126705, + 0.007843317463994026, + 0.005449241958558559, + 0.004478957504034042, + 0.07023798674345016, + 0.007730540819466114, + 0.05034932866692543, + 0.007384886033833027, + 0.009983294643461704, + 0.0075440905056893826, + 0.007082801777869463, + 0.013039955869317055, + 0.007478297688066959, + 0.010903410613536835, + 0.006043820176273584, + 0.0048482646234333515, + 0.005624421872198582, + 0.0051480066031217575, + 0.003738204948604107, + 0.005489904433488846, + 0.00402689166367054, + 0.0034585825633257627, + 0.005522773135453463, + 0.002914999146014452, + 0.0021515428088605404, + 0.00254045519977808, + 0.0012260772055014968, + 0.002233568113297224, + 0.001713928417302668, + 0.0014098850078880787, + 0.0023013241589069366, + 0.002158839488402009, + 0.001754780299961567, + 0.0015219207853078842, + 0.0018446188187226653, + 0.0017199525609612465, + 0.002063942840322852, + 0.0015809601172804832, + 0.0016452622367069125, + 0.007327589206397533, + 0.0043522873893380165, + 0.00722266873344779, + 0.0072178286500275135, + 0.0071522146463394165, + 0.007566298823803663, + 0.005322025157511234, + 0.006138178054243326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.03374454751610756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548891827464104 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012548891827464104, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548891827464104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19348548352718353, + 0.22451138496398926, + 0.06243816018104553, + 0.04119487851858139, + 0.15020348131656647, + 0.04960228130221367, + 0.026185566559433937, + 0.051307566463947296, + 0.10862167924642563, + 0.09244959056377411 + ], + "credit_map": [ + 0.027877410873770714, + 0.016089536249637604, + 0.020734457299113274, + 0.015812188386917114, + 0.051528867334127426, + 0.021240323781967163, + 0.020519351586699486, + 0.019683334976434708, + 0.06359267234802246, + 0.02149045653641224, + 0.014841617085039616, + 0.013179964385926723, + 0.015438645146787167, + 0.05980685353279114, + 0.012082478031516075, + 0.02407868206501007, + 0.009813244454562664, + 0.004981225356459618, + 0.009166562929749489, + 0.005467560142278671, + 0.011292721144855022, + 0.007378751877695322, + 0.00691866222769022, + 0.00741943484172225, + 0.006971936672925949, + 0.004627292975783348, + 0.004038435406982899, + 0.003044889308512211, + 0.005422665271908045, + 0.005731045734137297, + 0.00334432115778327, + 0.00801429059356451, + 0.015248370356857777, + 0.0056052436120808125, + 0.006955901626497507, + 0.004736399743705988, + 0.005167391616851091, + 0.061669863760471344, + 0.006887627299875021, + 0.04393269121646881, + 0.008714163675904274, + 0.0068358308635652065, + 0.005945670884102583, + 0.004099687095731497, + 0.008929584175348282, + 0.004746267572045326, + 0.005658715032041073, + 0.004672365263104439, + 0.003045780351385474, + 0.0031473401468247175, + 0.003557216376066208, + 0.0031131363939493895, + 0.004620871506631374, + 0.003115043044090271, + 0.002726871520280838, + 0.002859307685866952, + 0.00654799398034811, + 0.005387416575103998, + 0.009099735878407955, + 0.004955303389579058, + 0.00911443866789341, + 0.004324438516050577, + 0.005399059969931841, + 0.00647917902097106, + 0.010052641853690147, + 0.011952066794037819, + 0.012587182223796844, + 0.012290777638554573, + 0.027950536459684372, + 0.012423625215888023, + 0.009781374596059322, + 0.01158346701413393, + 0.017408980056643486, + 0.009130275808274746, + 0.011161893606185913, + 0.005959651432931423, + 0.010899710468947887, + 0.019147519022226334, + 0.0066205523908138275, + 0.012121007777750492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.012548891827464104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291800171136856 + ], + "opposite_logprob_tokens": [ + -3.773291826248169 + ], + "logprob_reward": -0.023291800171136856, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291800171136856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1279623657464981, + 0.11676319688558578, + 0.05910768732428551, + 0.05956077575683594, + 0.21226085722446442, + 0.080320805311203, + 0.11842815577983856, + 0.07340206205844879, + 0.07484231889247894, + 0.07735174894332886 + ], + "credit_map": [ + 0.029548345133662224, + 0.015632903203368187, + 0.008423835039138794, + 0.009915732778608799, + 0.022718943655490875, + 0.015915686264634132, + 0.009704414755105972, + 0.016102517023682594, + 0.0206894613802433, + 0.014292778447270393, + 0.021672386676073074, + 0.008570816367864609, + 0.009345979429781437, + 0.015019353479146957, + 0.00888784695416689, + 0.018284577876329422, + 0.012524722144007683, + 0.004710332024842501, + 0.0043353913351893425, + 0.004650511778891087, + 0.009650405496358871, + 0.007635531481355429, + 0.00448435777798295, + 0.011116436682641506, + 0.009887710213661194, + 0.005587107501924038, + 0.006133398972451687, + 0.004328879062086344, + 0.006523262243717909, + 0.009126572869718075, + 0.005185307469218969, + 0.01278853788971901, + 0.021341273561120033, + 0.006732163485139608, + 0.009187988005578518, + 0.006401958409696817, + 0.005105877760797739, + 0.08995604515075684, + 0.009368283674120903, + 0.06416726857423782, + 0.00860631838440895, + 0.009166629984974861, + 0.010307726450264454, + 0.006197749171406031, + 0.023186245933175087, + 0.0086665702983737, + 0.007213742006570101, + 0.00697582820430398, + 0.018443336710333824, + 0.01964266039431095, + 0.012314465828239918, + 0.010175934992730618, + 0.013814320787787437, + 0.017491662874817848, + 0.008449946530163288, + 0.018095824867486954, + 0.010875923559069633, + 0.008976073935627937, + 0.01593646965920925, + 0.0066849831491708755, + 0.007392475381493568, + 0.006475330330431461, + 0.0065951868891716, + 0.010465622879564762, + 0.008117490448057652, + 0.009259436279535294, + 0.00793385598808527, + 0.008590903133153915, + 0.01661134511232376, + 0.009269120171666145, + 0.007350169122219086, + 0.007709991652518511, + 0.01263525802642107, + 0.009301841259002686, + 0.009554216638207436, + 0.00670350668951869, + 0.007919668219983578, + 0.01380820944905281, + 0.007031864952296019, + 0.010397179052233696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.023291800171136856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986714839935303 + ], + "opposite_logprob_tokens": [ + -1.804867148399353 + ], + "logprob_reward": -0.17986714839935303, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986714839935303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09466443955898285, + 0.08093442767858505, + 0.1226484477519989, + 0.02475360408425331, + 0.100191131234169, + 0.15469692647457123, + 0.12532678246498108, + 0.05650776997208595, + 0.09065644443035126, + 0.1496199518442154 + ], + "credit_map": [ + 0.01300087384879589, + 0.009295357391238213, + 0.01743169315159321, + 0.008726323023438454, + 0.013656198978424072, + 0.012759869918227196, + 0.008297611027956009, + 0.011496512219309807, + 0.023382455110549927, + 0.00645901495590806, + 0.005786942783743143, + 0.003997381776571274, + 0.008360364474356174, + 0.020679563283920288, + 0.00403385516256094, + 0.00823485292494297, + 0.01702605001628399, + 0.009103096090257168, + 0.01865709386765957, + 0.013095586560666561, + 0.021813856437802315, + 0.012915382161736488, + 0.014795095659792423, + 0.015242298133671284, + 0.0040074298158288, + 0.002316331258043647, + 0.002936594421043992, + 0.0017669624648988247, + 0.0029723593033850193, + 0.003593391040340066, + 0.002124852268025279, + 0.005035684909671545, + 0.010113781318068504, + 0.0032388526014983654, + 0.0038440206553786993, + 0.003603963879868388, + 0.0027705244719982147, + 0.0415937677025795, + 0.004564900882542133, + 0.030461322516202927, + 0.016660969704389572, + 0.02271958440542221, + 0.017852740362286568, + 0.017613880336284637, + 0.02116125449538231, + 0.01791461743414402, + 0.027082020416855812, + 0.013691878877580166, + 0.013159355148673058, + 0.023614060133695602, + 0.013590514659881592, + 0.010834351181983948, + 0.017602497711777687, + 0.01705789379775524, + 0.009248986840248108, + 0.02021913416683674, + 0.008299813605844975, + 0.006319223437458277, + 0.012400872074067593, + 0.005762751214206219, + 0.006119659636169672, + 0.005250013433396816, + 0.004434439819306135, + 0.007921000942587852, + 0.008544879965484142, + 0.00977126881480217, + 0.01297619380056858, + 0.01051126979291439, + 0.018120724707841873, + 0.010526073165237904, + 0.009272805415093899, + 0.010933233425021172, + 0.03176435828208923, + 0.014500872232019901, + 0.017586547881364822, + 0.00839131511747837, + 0.016328517347574234, + 0.03291672468185425, + 0.009033365175127983, + 0.019098270684480667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.17986714839935303 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic overhead shot, beginning with a close-up of the eagle and smoothly zooming out as it moves, captures a majestic eagle perched atop the desolate, windswept summit of a towering mountain. Its snow-white head feathers, piercing, focused golden eyes, and a massive, hooked yellow beak gleam with a cool sheen in the first golden rays of morning sunlight. Its deep brown body is layered with feathers, and its powerful yellow talons grip a ridged, granite rock, shrouded in lingering snow. Behind it, a vast panorama of jagged, snow-capped peaks stretches to the horizon, while wisps of morning mist curl through the valley. After a moment of stillness, the eagle's body suddenly tensed, and then with an explosive yet extremely calm movement, it spread its wide wings steadily and powerfully. The camera then zoomed out to fully display its amazing wingspan. Each main flight feather was outlined with a golden edge in the backlight, as if declaring its sovereignty over the sky. The whole picture was filled with a tranquil atmosphere of majesty, freedom and untamed natural power.", + "seed": 4790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_11/index_6739_seed_4790_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": -0.11339629851281643, + "vlm_logprob": -0.11339629851281643, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.11339629851281643, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20528046786785126 + ], + "opposite_logprob_tokens": [ + -1.7052804231643677 + ], + "logprob_reward": -0.20528046786785126, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20528046786785126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06354966014623642, + 0.041059527546167374, + 0.18144665658473969, + 0.015955382958054543, + 0.36019375920295715, + 0.032563064247369766, + 0.02274216152727604, + 0.1913338601589203, + 0.05807497724890709, + 0.03308095782995224 + ], + "credit_map": [ + 0.01465883944183588, + 0.004822025075554848, + 0.005279710050672293, + 0.002968635642901063, + 0.006937946658581495, + 0.01972741261124611, + 0.003528621979057789, + 0.00562647171318531, + 0.008562294766306877, + 0.002627581125125289, + 0.003167974529787898, + 0.0014401818625628948, + 0.0021363424602895975, + 0.015245715156197548, + 0.0026742999907583, + 0.005205139517784119, + 0.04862635210156441, + 0.006136683281511068, + 0.011522084474563599, + 0.010592098347842693, + 0.005928167141973972, + 0.0793725848197937, + 0.011444681324064732, + 0.00782400369644165, + 0.003065136494114995, + 0.0011309351539239287, + 0.0020629102364182472, + 0.0010523982346057892, + 0.001710967393592, + 0.004599657375365496, + 0.001049173646606505, + 0.0012842054711654782, + 0.11277987062931061, + 0.01414159219712019, + 0.1385003924369812, + 0.020403215661644936, + 0.010716862045228481, + 0.026626184582710266, + 0.011892649345099926, + 0.02513299323618412, + 0.0086258789524436, + 0.004250459838658571, + 0.0037350179627537727, + 0.002620970830321312, + 0.00429430091753602, + 0.004405507352203131, + 0.0018639176851138473, + 0.0027670119889080524, + 0.002151465741917491, + 0.0013032930437475443, + 0.0020190721843391657, + 0.006170247215777636, + 0.0010491486173123121, + 0.0072798714973032475, + 0.001138062565587461, + 0.0016309998463839293, + 0.04511178657412529, + 0.01388969924300909, + 0.02223842591047287, + 0.06302841752767563, + 0.009908775798976421, + 0.012709190137684345, + 0.011247923597693443, + 0.0131996376439929, + 0.008543056435883045, + 0.01064788457006216, + 0.004696278367191553, + 0.008390836417675018, + 0.007156058214604855, + 0.008073005825281143, + 0.005477310158312321, + 0.005090547725558281, + 0.004826804623007774, + 0.004539793822914362, + 0.004293608479201794, + 0.0034901960752904415, + 0.0034600745420902967, + 0.004760457202792168, + 0.003112535923719406, + 0.004597487859427929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.20528046786785126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559369206428528 + ], + "opposite_logprob_tokens": [ + -4.387559413909912 + ], + "logprob_reward": -0.012559369206428528, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559369206428528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1554763913154602, + 0.14045125246047974, + 0.19207298755645752, + 0.046518780291080475, + 0.1809634417295456, + 0.013739170506596565, + 0.0386027954518795, + 0.16212612390518188, + 0.020578574389219284, + 0.04947052150964737 + ], + "credit_map": [ + 0.040380239486694336, + 0.0073346528224647045, + 0.011006327345967293, + 0.00610982533544302, + 0.01021268218755722, + 0.062273211777210236, + 0.008746201172471046, + 0.009413243271410465, + 0.030195528641343117, + 0.008856326341629028, + 0.010660400614142418, + 0.005110627505928278, + 0.0069684842601418495, + 0.05393523350358009, + 0.008897372521460056, + 0.01582726091146469, + 0.051497865468263626, + 0.006844743620604277, + 0.012688031420111656, + 0.010452241636812687, + 0.006772842723876238, + 0.08286309987306595, + 0.012098206207156181, + 0.008855936117470264, + 0.012754226103425026, + 0.0025520578492432833, + 0.0032365601509809494, + 0.0028126435354351997, + 0.002268965123221278, + 0.019057461991906166, + 0.0017479259986430407, + 0.0020889360457658768, + 0.055937010794878006, + 0.007445253431797028, + 0.06894657760858536, + 0.010322494432330132, + 0.005619200877845287, + 0.013611363247036934, + 0.006044591777026653, + 0.013036929070949554, + 0.0033049306366592646, + 0.0018884115852415562, + 0.0015170667320489883, + 0.0010981867089867592, + 0.0018031895160675049, + 0.0020268752705305815, + 0.0008328061085194349, + 0.0012677027843892574, + 0.004794157110154629, + 0.0030341586098074913, + 0.0036397746298462152, + 0.005858328193426132, + 0.002208765596151352, + 0.014047265984117985, + 0.0018116739811375737, + 0.0032086684368550777, + 0.04059122875332832, + 0.011677287518978119, + 0.017223183065652847, + 0.05354685336351395, + 0.00827058032155037, + 0.010317791253328323, + 0.009460195899009705, + 0.011038974858820438, + 0.0028573551680892706, + 0.0030992196407169104, + 0.0019966554827988148, + 0.003329529194161296, + 0.0025523288641124964, + 0.0029856220353394747, + 0.0019269944168627262, + 0.0018308676080778241, + 0.007115290500223637, + 0.006692572962492704, + 0.006272686179727316, + 0.00567716546356678, + 0.00570903904736042, + 0.006610440090298653, + 0.004014559090137482, + 0.007378761190921068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.012559369206428528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586861297488213 + ], + "opposite_logprob_tokens": [ + -3.8955869674682617 + ], + "logprob_reward": -0.020586861297488213, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586861297488213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18950195610523224, + 0.11923825740814209, + 0.10216968506574631, + 0.060132358223199844, + 0.1435866504907608, + 0.019238131120800972, + 0.11836318671703339, + 0.12062469869852066, + 0.0776338204741478, + 0.049511272460222244 + ], + "credit_map": [ + 0.05001463368535042, + 0.01112962793558836, + 0.014257535338401794, + 0.004628037568181753, + 0.01446380652487278, + 0.07508347928524017, + 0.01003015786409378, + 0.009894687682390213, + 0.027984637767076492, + 0.007015102077275515, + 0.009049639105796814, + 0.00484045036137104, + 0.007642890326678753, + 0.04911864176392555, + 0.007114761043339968, + 0.006472140550613403, + 0.029249828308820724, + 0.003699657740071416, + 0.006152746733278036, + 0.0044508627615869045, + 0.003416364314034581, + 0.04513940587639809, + 0.006475447677075863, + 0.003585370257496834, + 0.01615196093916893, + 0.0037983525544404984, + 0.0041756536811590195, + 0.003004414262250066, + 0.004242050461471081, + 0.023131227120757103, + 0.0025194077752530575, + 0.0031092860735952854, + 0.03875444456934929, + 0.016554584726691246, + 0.03834187984466553, + 0.008485323749482632, + 0.010998020879924297, + 0.011886544525623322, + 0.009940782561898232, + 0.008625073358416557, + 0.002643785672262311, + 0.003483717329800129, + 0.002259779255837202, + 0.0015676973853260279, + 0.002332554431632161, + 0.003090013051405549, + 0.0018423324218019843, + 0.002018251456320286, + 0.012820533476769924, + 0.005622791592031717, + 0.010227066464722157, + 0.027582963928580284, + 0.0049786376766860485, + 0.043750032782554626, + 0.004825958516448736, + 0.008555205538868904, + 0.02444254793226719, + 0.010186854749917984, + 0.016455532982945442, + 0.036606281995773315, + 0.007847176864743233, + 0.01013101264834404, + 0.007033035159111023, + 0.007922249846160412, + 0.012180520221590996, + 0.01132045965641737, + 0.006971242371946573, + 0.012598877772688866, + 0.008835837244987488, + 0.007952190935611725, + 0.01075427234172821, + 0.007020424120128155, + 0.0045285713858902454, + 0.006213789340108633, + 0.0045556980185210705, + 0.0047727725468575954, + 0.01063921581953764, + 0.007756347302347422, + 0.006487495731562376, + 0.004557380918413401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.020586861297488213 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270226538181305 + ], + "opposite_logprob_tokens": [ + -2.1270227432250977 + ], + "logprob_reward": -0.1270226538181305, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270226538181305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2410348653793335, + 0.026465347036719322, + 0.07010212540626526, + 0.028692614287137985, + 0.07497405260801315, + 0.02251540496945381, + 0.08641057461500168, + 0.3024919629096985, + 0.11819975823163986, + 0.0291132889688015 + ], + "credit_map": [ + 0.06866778433322906, + 0.009576483629643917, + 0.01682440936565399, + 0.0047388202510774136, + 0.012205216102302074, + 0.10375908017158508, + 0.012785608880221844, + 0.012477467767894268, + 0.004508734215050936, + 0.002672211965546012, + 0.001993877347558737, + 0.0012162065831944346, + 0.0027433650102466345, + 0.0072030313313007355, + 0.002094264840707183, + 0.004033655393868685, + 0.016900209710001945, + 0.0032836562022566795, + 0.004375616554170847, + 0.004925049841403961, + 0.0036575153935700655, + 0.02847628854215145, + 0.004528799094259739, + 0.003954993095248938, + 0.008710123598575592, + 0.0011731897247955203, + 0.002081851242110133, + 0.0011505532311275601, + 0.0012889057397842407, + 0.011857804842293262, + 0.0012375985970720649, + 0.0011925867293030024, + 0.02225515805184841, + 0.004731479100883007, + 0.025543246418237686, + 0.004549280274659395, + 0.0034321676939725876, + 0.006117385346442461, + 0.0032127443701028824, + 0.00513259693980217, + 0.0034605206456035376, + 0.003981905523687601, + 0.002742862096056342, + 0.0019242007983848453, + 0.0024277332704514265, + 0.003408142365515232, + 0.001942995353601873, + 0.002627046313136816, + 0.01132359728217125, + 0.004653966519981623, + 0.006851089186966419, + 0.01718040369451046, + 0.003975851461291313, + 0.032089971005916595, + 0.003326274687424302, + 0.007009420078247786, + 0.06309515982866287, + 0.024883996695280075, + 0.040462687611579895, + 0.09220189601182938, + 0.019102642312645912, + 0.023975007236003876, + 0.017964789643883705, + 0.020805800333619118, + 0.01619020290672779, + 0.018920594826340675, + 0.01160390954464674, + 0.017629528418183327, + 0.013211103156208992, + 0.012660598382353783, + 0.01839366927742958, + 0.009590153582394123, + 0.0032748838420957327, + 0.004380645230412483, + 0.002625759458169341, + 0.0036300725769251585, + 0.003402560018002987, + 0.0030315069016069174, + 0.006081660743802786, + 0.0026862008962780237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.1270226538181305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153214037418365 + ], + "opposite_logprob_tokens": [ + -1.7015321254730225 + ], + "logprob_reward": -0.20153214037418365, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153214037418365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33876970410346985, + 0.019217027351260185, + 0.049240536987781525, + 0.031783849000930786, + 0.10283386707305908, + 0.03371339291334152, + 0.17937687039375305, + 0.11750989407300949, + 0.08253507316112518, + 0.04501982778310776 + ], + "credit_map": [ + 0.09507910162210464, + 0.013414044864475727, + 0.024110760539770126, + 0.00764263654127717, + 0.01764179766178131, + 0.1424330174922943, + 0.01750456728041172, + 0.020943772047758102, + 0.0025025433860719204, + 0.001658533001318574, + 0.0024997182190418243, + 0.0012820016127079725, + 0.003417222760617733, + 0.002727407729253173, + 0.0018754046177491546, + 0.003254195209592581, + 0.006926303263753653, + 0.005663011223077774, + 0.00400113919749856, + 0.006897857412695885, + 0.004750038497149944, + 0.014308659359812737, + 0.003073036437854171, + 0.0036204904317855835, + 0.00754533801227808, + 0.0021188033279031515, + 0.002631781855598092, + 0.0014999101404100657, + 0.004314499907195568, + 0.009594622999429703, + 0.002248009666800499, + 0.0018308807630091906, + 0.022590067237615585, + 0.004877730272710323, + 0.04125073924660683, + 0.008535015396773815, + 0.0048052347265183926, + 0.007002629339694977, + 0.004072990268468857, + 0.009699457325041294, + 0.0060385228134691715, + 0.005671248305588961, + 0.004583751317113638, + 0.0026401574723422527, + 0.0039833663031458855, + 0.004660374950617552, + 0.0023271404206752777, + 0.0038088341243565083, + 0.020689111202955246, + 0.009009876288473606, + 0.014805211685597897, + 0.04039223492145538, + 0.007689303252846003, + 0.06510978192090988, + 0.007504502777010202, + 0.014176844619214535, + 0.013060538098216057, + 0.008208009414374828, + 0.025948025286197662, + 0.030439505353569984, + 0.008808808401226997, + 0.014981645159423351, + 0.00681642210111022, + 0.009246936067938805, + 0.01048357505351305, + 0.014040401205420494, + 0.007759166415780783, + 0.01257074810564518, + 0.009202178567647934, + 0.00919393915683031, + 0.012207592837512493, + 0.0070774792693555355, + 0.00692920433357358, + 0.006730204913765192, + 0.004449218977242708, + 0.004723510704934597, + 0.004768774379044771, + 0.00695494981482625, + 0.005824711173772812, + 0.004639253485947847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.20153214037418365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic overhead shot, beginning with a close-up of the eagle and smoothly zooming out as it moves, captures a majestic eagle perched atop the desolate, windswept summit of a towering mountain. Its snow-white head feathers, piercing, focused golden eyes, and a massive, hooked yellow beak gleam with a cool sheen in the first golden rays of morning sunlight. Its deep brown body is layered with feathers, and its powerful yellow talons grip a ridged, granite rock, shrouded in lingering snow. Behind it, a vast panorama of jagged, snow-capped peaks stretches to the horizon, while wisps of morning mist curl through the valley. After a moment of stillness, the eagle's body suddenly tensed, and then with an explosive yet extremely calm movement, it spread its wide wings steadily and powerfully. The camera then zoomed out to fully display its amazing wingspan. Each main flight feather was outlined with a golden edge in the backlight, as if declaring its sovereignty over the sky. The whole picture was filled with a tranquil atmosphere of majesty, freedom and untamed natural power.", + "seed": 4791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_11/index_6739_seed_4791_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": -0.10055454652756453, + "vlm_logprob": -0.10055454652756453, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.10055454652756453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09300202876329422 + ], + "opposite_logprob_tokens": [ + -2.4680020809173584 + ], + "logprob_reward": -0.09300202876329422, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09300202876329422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09819556027650833, + 0.23288895189762115, + 0.02930683270096779, + 0.10689394176006317, + 0.0577329657971859, + 0.1916016787290573, + 0.07606389373540878, + 0.16812120378017426, + 0.015319171361625195, + 0.02387586422264576 + ], + "credit_map": [ + 0.020560506731271744, + 0.012302986346185207, + 0.008588576689362526, + 0.007551169954240322, + 0.010374890640377998, + 0.01546345092356205, + 0.007804353721439838, + 0.01554962620139122, + 0.022852689027786255, + 0.020120825618505478, + 0.016693640500307083, + 0.01491572242230177, + 0.016519421711564064, + 0.07932501286268234, + 0.014805278740823269, + 0.0476563423871994, + 0.005279000382870436, + 0.003736283630132675, + 0.0031122523359954357, + 0.0018774353666231036, + 0.0029856639448553324, + 0.006410126108676195, + 0.002063605235889554, + 0.0038424672093242407, + 0.003953786566853523, + 0.01220263447612524, + 0.005089328158646822, + 0.0029919040389358997, + 0.004144302569329739, + 0.04643253982067108, + 0.003029401181265712, + 0.02905004471540451, + 0.00797684583812952, + 0.007416099775582552, + 0.00632353825494647, + 0.0068833776749670506, + 0.00657148752361536, + 0.011430089361965656, + 0.004569668788462877, + 0.0065618581138551235, + 0.009998099878430367, + 0.02290641888976097, + 0.008691094815731049, + 0.00540356757119298, + 0.009449033066630363, + 0.08202783763408661, + 0.00480191083624959, + 0.04832371696829796, + 0.0036526855546981096, + 0.008653335273265839, + 0.0036138540599495173, + 0.002052617957815528, + 0.004219772294163704, + 0.03387966379523277, + 0.0020957475062459707, + 0.017896223813295364, + 0.02406686171889305, + 0.021407026797533035, + 0.02346542477607727, + 0.023949284106492996, + 0.011177252978086472, + 0.03051448054611683, + 0.016280261799693108, + 0.01726062037050724, + 0.0020477052312344313, + 0.0021189628168940544, + 0.0015071001835167408, + 0.0012251485604792833, + 0.002095168689265847, + 0.0040610674768686295, + 0.0009510608506388962, + 0.0013129577273502946, + 0.002739465329796076, + 0.006654142867773771, + 0.0018444492015987635, + 0.0023221378214657307, + 0.0019193688640370965, + 0.002825787290930748, + 0.001785597880370915, + 0.0037849151995033026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.09300202876329422 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209638349711895 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209638349711895, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209638349711895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03405367210507393, + 0.21274511516094208, + 0.027864128351211548, + 0.07389681041240692, + 0.07141389697790146, + 0.3054456114768982, + 0.033369068056344986, + 0.19748744368553162, + 0.019254663959145546, + 0.024469643831253052 + ], + "credit_map": [ + 0.004853246733546257, + 0.003920364659279585, + 0.0033410508185625076, + 0.0030976561829447746, + 0.005461118649691343, + 0.0050800871104002, + 0.0041131130419671535, + 0.004187031649053097, + 0.019932901486754417, + 0.019223595038056374, + 0.0144491633400321, + 0.013780389912426472, + 0.01276986300945282, + 0.07279210537672043, + 0.014214781112968922, + 0.045582301914691925, + 0.004988843109458685, + 0.0030273732263594866, + 0.003232011804357171, + 0.0019113827729597688, + 0.0027819236274808645, + 0.0060692476108670235, + 0.002329518087208271, + 0.00352382636629045, + 0.0033187265507876873, + 0.007969330064952374, + 0.0039484454318881035, + 0.0032613351941108704, + 0.0034405544865876436, + 0.030337294563651085, + 0.0028314776718616486, + 0.01878964528441429, + 0.00894925557076931, + 0.00850129034370184, + 0.00854092463850975, + 0.008788088336586952, + 0.007901170291006565, + 0.014175968244671822, + 0.006093581207096577, + 0.008463611826300621, + 0.015849636867642403, + 0.036479875445365906, + 0.013618401251733303, + 0.00858649518340826, + 0.014575710520148277, + 0.1315658837556839, + 0.007470312062650919, + 0.07729925215244293, + 0.0037055357825011015, + 0.003677561180666089, + 0.004027059301733971, + 0.002426385646685958, + 0.002773548476397991, + 0.01042097620666027, + 0.0022004488855600357, + 0.004137550015002489, + 0.023533303290605545, + 0.02075234241783619, + 0.03098505735397339, + 0.02752947434782982, + 0.013932652771472931, + 0.03395634517073631, + 0.0236839447170496, + 0.02311428263783455, + 0.0022439779713749886, + 0.0033775405026972294, + 0.0017689315136522055, + 0.0018987994408234954, + 0.0034426203928887844, + 0.002772017614915967, + 0.0017384913517162204, + 0.002012283308431506, + 0.0033547282218933105, + 0.0030002943240106106, + 0.0033816725481301546, + 0.002392173744738102, + 0.002806486329063773, + 0.003905727993696928, + 0.002830272074788809, + 0.0027982862666249275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.014209638349711895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196004718542099 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196004718542099, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196004718542099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12941408157348633, + 0.05417751893401146, + 0.0320744626224041, + 0.036303967237472534, + 0.02263035997748375, + 0.1989155262708664, + 0.2347133904695511, + 0.0633285865187645, + 0.03696345165371895, + 0.1914786845445633 + ], + "credit_map": [ + 0.015865860506892204, + 0.0165143683552742, + 0.010972063057124615, + 0.011629012413322926, + 0.024455223232507706, + 0.017211008816957474, + 0.013997286558151245, + 0.018769264221191406, + 0.005441978108137846, + 0.008046652190387249, + 0.0039646984077990055, + 0.004683167673647404, + 0.008983018808066845, + 0.007514762692153454, + 0.007285505998879671, + 0.008257733657956123, + 0.0038845089729875326, + 0.003382645081728697, + 0.002097747288644314, + 0.002077277982607484, + 0.008998856879770756, + 0.005737352184951305, + 0.0033573347609490156, + 0.0025387408677488565, + 0.0022772701922804117, + 0.004583882167935371, + 0.002001892775297165, + 0.001813409966416657, + 0.005417302716523409, + 0.011097414419054985, + 0.0029874593019485474, + 0.006125336512923241, + 0.0019489203114062548, + 0.0026107628364115953, + 0.002386644948273897, + 0.003956452012062073, + 0.0028013326227664948, + 0.00340760825201869, + 0.003626524470746517, + 0.0018921144073829055, + 0.009623009711503983, + 0.02316092513501644, + 0.008874663151800632, + 0.005258592311292887, + 0.00952725950628519, + 0.08613594621419907, + 0.005912172142416239, + 0.05042295530438423, + 0.009406683035194874, + 0.025702370330691338, + 0.008755779825150967, + 0.006948363501578569, + 0.013190806843340397, + 0.10483086109161377, + 0.0073439027182757854, + 0.05853462964296341, + 0.009042317047715187, + 0.008090732619166374, + 0.009512045420706272, + 0.006837988272309303, + 0.006665828637778759, + 0.008302975445985794, + 0.007682980503886938, + 0.007193718105554581, + 0.004780094604939222, + 0.005160869564861059, + 0.0037562656216323376, + 0.0041899005882442, + 0.005310576409101486, + 0.005262291990220547, + 0.00367019884288311, + 0.00483325868844986, + 0.017668118700385094, + 0.05720379948616028, + 0.011775506660342216, + 0.012811446562409401, + 0.01814708299934864, + 0.025192881003022194, + 0.014145223423838615, + 0.03453460708260536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.014196004718542099 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985281348228455 + ], + "opposite_logprob_tokens": [ + -1.804852843284607 + ], + "logprob_reward": -0.17985281348228455, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985281348228455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04609253630042076, + 0.052705928683280945, + 0.022614475339651108, + 0.03158942982554436, + 0.014462057501077652, + 0.11390476673841476, + 0.27367469668388367, + 0.02934269979596138, + 0.03467920795083046, + 0.3809341788291931 + ], + "credit_map": [ + 0.01205514557659626, + 0.00511827040463686, + 0.004667687229812145, + 0.003846710082143545, + 0.003957572393119335, + 0.0053580766543745995, + 0.004124272149056196, + 0.006964799948036671, + 0.006677351426333189, + 0.0076782358810305595, + 0.002930613234639168, + 0.003343723015859723, + 0.0043957908637821674, + 0.016771020367741585, + 0.0027488605119287968, + 0.008160335943102837, + 0.0036482885479927063, + 0.002961261197924614, + 0.0017795165767893195, + 0.0016450817929580808, + 0.004423143342137337, + 0.003331627929583192, + 0.0030703290831297636, + 0.0017552272183820605, + 0.0017370152054354548, + 0.0027498216368258, + 0.0015897738048806787, + 0.0012402242282405496, + 0.0035911279264837503, + 0.010759811848402023, + 0.0030687805265188217, + 0.006852875929325819, + 0.0013634057249873877, + 0.0016224745195358992, + 0.0017324549844488502, + 0.002626359462738037, + 0.0019052724819630384, + 0.0019123960519209504, + 0.0014617093838751316, + 0.0018379840767011046, + 0.005742567125707865, + 0.013421757146716118, + 0.00482026906684041, + 0.003319455310702324, + 0.006496307905763388, + 0.04799201712012291, + 0.0038175873924046755, + 0.02829480916261673, + 0.010640781372785568, + 0.03106795996427536, + 0.010386641137301922, + 0.006979181431233883, + 0.014744885265827179, + 0.12494666874408722, + 0.006056100595742464, + 0.0688524842262268, + 0.003300907090306282, + 0.0038322927430272102, + 0.004780724178999662, + 0.0037155053578317165, + 0.003543758997693658, + 0.0034130404237657785, + 0.003891671309247613, + 0.0028647996950894594, + 0.0039940327405929565, + 0.00492031080648303, + 0.003960900008678436, + 0.0034463268239051104, + 0.006231033243238926, + 0.004768716637045145, + 0.0036281889770179987, + 0.0037296991795301437, + 0.03513915464282036, + 0.12412708252668381, + 0.022487754002213478, + 0.02539685368537903, + 0.032236430794000626, + 0.044718362390995026, + 0.026523608714342117, + 0.07030493021011353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.17985281348228455 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20151224732398987 + ], + "opposite_logprob_tokens": [ + -1.7015122175216675 + ], + "logprob_reward": -0.20151224732398987, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20151224732398987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07574116438627243, + 0.09122040867805481, + 0.034597042948007584, + 0.04060572385787964, + 0.036266352981328964, + 0.24653714895248413, + 0.30594488978385925, + 0.05329212546348572, + 0.047109074890613556, + 0.0686860904097557 + ], + "credit_map": [ + 0.015141718089580536, + 0.007705273572355509, + 0.006236330606043339, + 0.006195950787514448, + 0.01147571299225092, + 0.011911407113075256, + 0.007802438456565142, + 0.009272327646613121, + 0.012072044424712658, + 0.011077803559601307, + 0.00715143047273159, + 0.008885963819921017, + 0.01135307364165783, + 0.023172931745648384, + 0.008274491876363754, + 0.009232665412127972, + 0.0053512039594352245, + 0.003966530319303274, + 0.002425032202154398, + 0.0017236570129171014, + 0.006197480950504541, + 0.00791969895362854, + 0.0027213015127927065, + 0.00429213372990489, + 0.0018161461921408772, + 0.004247313365340233, + 0.0020398276392370462, + 0.0015323817497119308, + 0.0021825814619660378, + 0.016176113858819008, + 0.0020096690859645605, + 0.010601690970361233, + 0.0030074191745370626, + 0.004458559677004814, + 0.002678284887224436, + 0.00392884761095047, + 0.006997391581535339, + 0.005444509908556938, + 0.005739013198763132, + 0.004012325778603554, + 0.010130916722118855, + 0.02981613203883171, + 0.011577748693525791, + 0.006211213301867247, + 0.013852370902895927, + 0.10867160558700562, + 0.008771357126533985, + 0.05750580504536629, + 0.011944725178182125, + 0.03414304554462433, + 0.011441814713180065, + 0.008289401419460773, + 0.016444727778434753, + 0.13825847208499908, + 0.007240475621074438, + 0.07818222790956497, + 0.0073044379241764545, + 0.00692471070215106, + 0.006833892315626144, + 0.007364196702837944, + 0.0061452859081327915, + 0.006367927882820368, + 0.007009300868958235, + 0.005342372693121433, + 0.005416293628513813, + 0.008046511560678482, + 0.004688852000981569, + 0.006133701652288437, + 0.005538146011531353, + 0.008649000898003578, + 0.0040471842512488365, + 0.004589383024722338, + 0.006533022969961166, + 0.020167645066976547, + 0.004450567066669464, + 0.004307880066335201, + 0.007115784101188183, + 0.007972443476319313, + 0.005107370670884848, + 0.013031375594437122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.20151224732398987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic overhead shot, beginning with a close-up of the eagle and smoothly zooming out as it moves, captures a majestic eagle perched atop the desolate, windswept summit of a towering mountain. Its snow-white head feathers, piercing, focused golden eyes, and a massive, hooked yellow beak gleam with a cool sheen in the first golden rays of morning sunlight. Its deep brown body is layered with feathers, and its powerful yellow talons grip a ridged, granite rock, shrouded in lingering snow. Behind it, a vast panorama of jagged, snow-capped peaks stretches to the horizon, while wisps of morning mist curl through the valley. After a moment of stillness, the eagle's body suddenly tensed, and then with an explosive yet extremely calm movement, it spread its wide wings steadily and powerfully. The camera then zoomed out to fully display its amazing wingspan. Each main flight feather was outlined with a golden edge in the backlight, as if declaring its sovereignty over the sky. The whole picture was filled with a tranquil atmosphere of majesty, freedom and untamed natural power.", + "seed": 4792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_11/index_6739_seed_4792_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": -0.02900671907700598, + "vlm_logprob": -0.02900671907700598, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02900671907700598, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008939128369092941 + ], + "opposite_logprob_tokens": [ + -5.008939266204834 + ], + "logprob_reward": -0.008939128369092941, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008939128369092941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08607102930545807, + 0.33038246631622314, + 0.07189539819955826, + 0.04049861431121826, + 0.11364243924617767, + 0.06388822942972183, + 0.06494751572608948, + 0.0949498862028122, + 0.08081068098545074, + 0.052913788706064224 + ], + "credit_map": [ + 0.010672642849385738, + 0.011863039806485176, + 0.014911732636392117, + 0.0070522441528737545, + 0.009928284212946892, + 0.012406266294419765, + 0.009129803627729416, + 0.01010701060295105, + 0.02719011902809143, + 0.026831049472093582, + 0.09196693450212479, + 0.015165315009653568, + 0.011089954525232315, + 0.024571625515818596, + 0.017721060663461685, + 0.11584638804197311, + 0.011422696523368359, + 0.013780400156974792, + 0.005797309335321188, + 0.005790958646684885, + 0.009391297586262226, + 0.015129763633012772, + 0.005244845990091562, + 0.005338120739907026, + 0.005593206267803907, + 0.005710212513804436, + 0.0032726721838116646, + 0.0026792343705892563, + 0.005913905333727598, + 0.01085266750305891, + 0.0030763554386794567, + 0.0034003572072833776, + 0.0159249659627676, + 0.019102169200778008, + 0.018104610964655876, + 0.012853037565946579, + 0.013113916851580143, + 0.013427714817225933, + 0.009464144706726074, + 0.011651875451207161, + 0.010340359061956406, + 0.006401931867003441, + 0.007782614789903164, + 0.006323718465864658, + 0.005067923106253147, + 0.016290800645947456, + 0.0055796923115849495, + 0.006101180799305439, + 0.007124970201402903, + 0.007511259987950325, + 0.008882506750524044, + 0.006579121109098196, + 0.004173953086137772, + 0.019300809130072594, + 0.004980739206075668, + 0.0063941520638763905, + 0.009211801923811436, + 0.0204083863645792, + 0.00919431820511818, + 0.01124529168009758, + 0.0079422602429986, + 0.014764039777219296, + 0.0074191647581756115, + 0.014764618128538132, + 0.01606842130422592, + 0.007217738311737776, + 0.010811551474034786, + 0.005985934287309647, + 0.008867016062140465, + 0.006437183823436499, + 0.009287643246352673, + 0.01613517850637436, + 0.008360653184354305, + 0.006739297416061163, + 0.004270447418093681, + 0.004106841050088406, + 0.0073554255068302155, + 0.008058853447437286, + 0.00437474949285388, + 0.009647517465054989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.008939128369092941 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624300705268979 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.003624300705268979, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624300705268979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07837193459272385, + 0.3645714819431305, + 0.12785616517066956, + 0.05666195601224899, + 0.07146249711513519, + 0.03082325868308544, + 0.08342510461807251, + 0.08968538790941238, + 0.05421772971749306, + 0.04292447119951248 + ], + "credit_map": [ + 0.008882886730134487, + 0.010728112421929836, + 0.011649071238934994, + 0.006319418549537659, + 0.011032603681087494, + 0.008847047574818134, + 0.012009111233055592, + 0.0089036850258708, + 0.03415994346141815, + 0.03149024397134781, + 0.10301923006772995, + 0.016975760459899902, + 0.011003823019564152, + 0.026534661650657654, + 0.01848178170621395, + 0.12290602922439575, + 0.016913222149014473, + 0.019418984651565552, + 0.018369875848293304, + 0.01518016867339611, + 0.02121882699429989, + 0.012327861972153187, + 0.01253858208656311, + 0.011888639070093632, + 0.0064324550330638885, + 0.006600058637559414, + 0.005158939398825169, + 0.0041021849028766155, + 0.015220241621136665, + 0.007630649954080582, + 0.006405483931303024, + 0.005111941136419773, + 0.008431782945990562, + 0.008277815766632557, + 0.01030491292476654, + 0.007322288118302822, + 0.010419822297990322, + 0.0074033779092133045, + 0.00796587485820055, + 0.011336627416312695, + 0.003532376140356064, + 0.0035436314065009356, + 0.005162184126675129, + 0.003306661732494831, + 0.00311618042178452, + 0.00506216986104846, + 0.004322631284594536, + 0.002777424408122897, + 0.008763957768678665, + 0.006949480623006821, + 0.017562130466103554, + 0.008272428996860981, + 0.004896094091236591, + 0.028268327936530113, + 0.004265960771590471, + 0.004446721635758877, + 0.011055167764425278, + 0.01396742183715105, + 0.010870052501559258, + 0.00664997473359108, + 0.007678813301026821, + 0.011298386380076408, + 0.008483538404107094, + 0.019682036712765694, + 0.0068201301619410515, + 0.0047405436635017395, + 0.00944235734641552, + 0.004562956281006336, + 0.006217245012521744, + 0.004878317005932331, + 0.006258326582610607, + 0.011297853663563728, + 0.007835670374333858, + 0.005595149472355843, + 0.004815689288079739, + 0.0029584455769509077, + 0.006593652069568634, + 0.005778128281235695, + 0.003498125122860074, + 0.005849612411111593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.003624300705268979 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782234206795692 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782234206795692, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782234206795692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11465821415185928, + 0.09695613384246826, + 0.1798640489578247, + 0.033932045102119446, + 0.03802092373371124, + 0.09459619224071503, + 0.08853548020124435, + 0.18776679039001465, + 0.09025652706623077, + 0.07541362941265106 + ], + "credit_map": [ + 0.012666141614317894, + 0.01162401307374239, + 0.016939077526330948, + 0.010589215904474258, + 0.019351458176970482, + 0.012264406308531761, + 0.015779180452227592, + 0.015444719232618809, + 0.008415577001869678, + 0.008380967192351818, + 0.026054514572024345, + 0.004615982063114643, + 0.005600062198936939, + 0.006591239478439093, + 0.006149617489427328, + 0.031148172914981842, + 0.017943091690540314, + 0.023273183032870293, + 0.038510385900735855, + 0.010072545148432255, + 0.013528668321669102, + 0.051873523741960526, + 0.011182761751115322, + 0.013479888439178467, + 0.005462115164846182, + 0.004522989504039288, + 0.0040825712494552135, + 0.003045061370357871, + 0.006961285602301359, + 0.004138606134802103, + 0.0028298089746385813, + 0.002889609895646572, + 0.0051759653724730015, + 0.004595547914505005, + 0.004234533756971359, + 0.0035568983294069767, + 0.005449418444186449, + 0.005701588932424784, + 0.004727022722363472, + 0.004579948261380196, + 0.02207752875983715, + 0.009774760343134403, + 0.009636206552386284, + 0.007829662412405014, + 0.008996299467980862, + 0.021839220076799393, + 0.006393224000930786, + 0.008049285039305687, + 0.011606146581470966, + 0.00727922422811389, + 0.014781390316784382, + 0.0083804652094841, + 0.00715923635289073, + 0.0181095190346241, + 0.01323289517313242, + 0.007986605167388916, + 0.019144896417856216, + 0.03936508670449257, + 0.018424253910779953, + 0.016071844846010208, + 0.016194604337215424, + 0.026064371690154076, + 0.015231847763061523, + 0.037269894033670425, + 0.010527809150516987, + 0.01054304651916027, + 0.011881810612976551, + 0.0070325834676623344, + 0.011017820797860622, + 0.01224604807794094, + 0.01322034653276205, + 0.013787061907351017, + 0.012541809119284153, + 0.010020137764513493, + 0.008952331729233265, + 0.005715217906981707, + 0.012222083285450935, + 0.010005244985222816, + 0.006707473658025265, + 0.009249327704310417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.009782234206795692 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033683691173791885 + ], + "opposite_logprob_tokens": [ + -3.4086837768554688 + ], + "logprob_reward": -0.033683691173791885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033683691173791885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052455902099609375, + 0.08897508680820465, + 0.07208459824323654, + 0.03887244313955307, + 0.05875987932085991, + 0.1132373958826065, + 0.22219400107860565, + 0.09743325412273407, + 0.1809832602739334, + 0.07500416785478592 + ], + "credit_map": [ + 0.006349984556436539, + 0.008627338334918022, + 0.009042857214808464, + 0.00438371067866683, + 0.006052357144653797, + 0.005262608174234629, + 0.007579609286040068, + 0.00515744648873806, + 0.012591306120157242, + 0.018050601705908775, + 0.015174252912402153, + 0.005670131649821997, + 0.005253045354038477, + 0.009257313795387745, + 0.004782919306308031, + 0.01819552667438984, + 0.006956339813768864, + 0.011863892897963524, + 0.010879745706915855, + 0.006156373769044876, + 0.010343415662646294, + 0.013636489398777485, + 0.006005637347698212, + 0.006242707837373018, + 0.005508121568709612, + 0.003243746468797326, + 0.004682742990553379, + 0.001994258025661111, + 0.0056721423752605915, + 0.01207044068723917, + 0.0025902842171490192, + 0.003110710997134447, + 0.006615850608795881, + 0.004658122081309557, + 0.006178875919431448, + 0.005926411133259535, + 0.010731588117778301, + 0.011548316106200218, + 0.00606862548738718, + 0.007032101042568684, + 0.013569246046245098, + 0.010200618766248226, + 0.021295633167028427, + 0.01134361233562231, + 0.00781650934368372, + 0.03353070467710495, + 0.007441763766109943, + 0.008039310574531555, + 0.03202114999294281, + 0.017025986686348915, + 0.035679660737514496, + 0.020294861868023872, + 0.015806464478373528, + 0.047929972410202026, + 0.035001240670681, + 0.0184346754103899, + 0.009257245808839798, + 0.022882137447595596, + 0.011822571977972984, + 0.010744769126176834, + 0.009206226095557213, + 0.011287168599665165, + 0.009481270797550678, + 0.012751868925988674, + 0.028876785188913345, + 0.014918462373316288, + 0.023253893479704857, + 0.014157927595078945, + 0.02630516141653061, + 0.02540060132741928, + 0.01528897974640131, + 0.03278147801756859, + 0.010646063834428787, + 0.010489094071090221, + 0.021519852802157402, + 0.004665226675570011, + 0.008211720734834671, + 0.0061978530138731, + 0.0046632797457277775, + 0.00861108023673296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.033683691173791885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on the summit of a towering mountain with jagged, snow-capped peaks in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the eagle perched on a ridged, granite rock covered with lingering snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle have snow-white head feathers, piercing golden eyes, and a massive hooked yellow beak?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the eagle spread its wide wings with a steady, powerful motion after a moment of stillness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a close-up and smoothly zoom out to display the eagle's full wingspan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900424093008041 + ], + "opposite_logprob_tokens": [ + -2.4640042781829834 + ], + "logprob_reward": -0.08900424093008041, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900424093008041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057241760194301605, + 0.17107002437114716, + 0.11779800802469254, + 0.04318404942750931, + 0.09288561344146729, + 0.10606279969215393, + 0.09578296542167664, + 0.11075518280267715, + 0.10371682047843933, + 0.10150278359651566 + ], + "credit_map": [ + 0.008987278677523136, + 0.005683029070496559, + 0.007924029603600502, + 0.004224415868520737, + 0.01046384684741497, + 0.005901208147406578, + 0.006833567284047604, + 0.007224384229630232, + 0.016271237283945084, + 0.008120313286781311, + 0.038041651248931885, + 0.0101648373529315, + 0.006357200909405947, + 0.010938202030956745, + 0.011284371837973595, + 0.06989219784736633, + 0.012731016613543034, + 0.011776028200984001, + 0.01268052589148283, + 0.008911669254302979, + 0.01927672140300274, + 0.031037451699376106, + 0.010371851734817028, + 0.011012749746441841, + 0.005009779240936041, + 0.005295406095683575, + 0.006104990374296904, + 0.002435895148664713, + 0.005013591609895229, + 0.012932517565786839, + 0.002981080673635006, + 0.0034107889514416456, + 0.010315054096281528, + 0.010589398443698883, + 0.009034828282892704, + 0.007422496564686298, + 0.01674354448914528, + 0.017337165772914886, + 0.011097935028374195, + 0.010345187038183212, + 0.014047999866306782, + 0.011472233571112156, + 0.012511031702160835, + 0.011578315868973732, + 0.01126670092344284, + 0.02615388296544552, + 0.010553267784416676, + 0.008479364216327667, + 0.010821613483130932, + 0.012729361653327942, + 0.010840472765266895, + 0.008074704557657242, + 0.011679122224450111, + 0.02029699645936489, + 0.012517914175987244, + 0.008822784759104252, + 0.01094876416027546, + 0.015607827343046665, + 0.017455676570534706, + 0.00824886653572321, + 0.008647827431559563, + 0.015142777934670448, + 0.008784874342381954, + 0.025918571278452873, + 0.01745806820690632, + 0.011125865392386913, + 0.012191103771328926, + 0.008068901486694813, + 0.014706035144627094, + 0.012052631936967373, + 0.016121212393045425, + 0.011993001215159893, + 0.012586980126798153, + 0.006717935670167208, + 0.01816585846245289, + 0.010579897090792656, + 0.011117316782474518, + 0.012010214850306511, + 0.007885012775659561, + 0.022439561784267426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.08900424093008041 + } + ], + "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/0000131/group_12/index_tensor([6526, 6526])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_12/index_tensor([6526, 6526])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..310693c027aac896b63d6aa9937b6d9b8beae601 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_12/index_tensor([6526, 6526])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot, shot from a level angle, focuses on a weary female researcher seated at a cluttered lab table covered with lab equipment. Her dark brown hair is casually tied into a messy bun, a few strands clinging to her cheeks with sweat. A pair of clear goggles are pushed up on her forehead, revealing her sunken eyes and faint dark circles. She wears a slightly wrinkled white lab coat, the collar of which peeks out from under a gray crew-neck T-shirt. Before her, a row of glass beakers and Erlenmeyer flasks hold liquids of various colors—one a vibrant turquoise, another a deep amber—several of which emit a constant, fine bubble and a soft hissing sound. In the background, the cold white light of overhead fluorescent lights fills the room, casting harsh reflections on the stainless steel equipment and glassware. Her shoulders are slightly hunched forward from prolonged concentration, and her gaze, initially blank, stares at the bubbling liquid before her, but her eyelids grow increasingly heavy. Finally, she seemed to give up resistance and slowly, almost reverently, closed her eyes. Her tense jawline finally relaxed at this moment, and she was immersed in a brief but profound tranquility, which contrasted sharply with the ongoing chemical reaction around her, creating an atmosphere of concentration and human vulnerability intertwined between scientific rigor and human exhaustion.", + "seed": 17396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_12/index_6526_seed_17396_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": -0.04256161991506815, + "vlm_logprob": -0.04256161991506815, + "vlm_yes_no_margin": 4.1999999523162845, + "vlm_token_credit": -0.04256161991506815, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.013290622271597385 + ], + "opposite_logprob_tokens": [ + -6.1382904052734375 + ], + "logprob_reward": -0.013290622271597385, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.013290622271597385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17114602029323578, + 0.09775281697511673, + 0.03210366889834404, + 0.07624559104442596, + 0.060311608016490936, + 0.048315584659576416, + 0.10751111060380936, + 0.05788847431540489, + 0.1659204214811325, + 0.18280471861362457 + ], + "credit_map": [ + 0.02715362422168255, + 0.019277984276413918, + 0.02557266131043434, + 0.024494342505931854, + 0.013024043291807175, + 0.018970061093568802, + 0.02402419224381447, + 0.018629103899002075, + 0.01305360347032547, + 0.009626684710383415, + 0.01325695589184761, + 0.008135490119457245, + 0.007567716296762228, + 0.010011074133217335, + 0.02369767241179943, + 0.012403621338307858, + 0.00430332263931632, + 0.003617905778810382, + 0.004478538874536753, + 0.0028251069597899914, + 0.0023572896607220173, + 0.004410198889672756, + 0.004751094151288271, + 0.005360212177038193, + 0.012668805196881294, + 0.004795408342033625, + 0.008673080243170261, + 0.004882958251982927, + 0.0053810784593224525, + 0.00787714309990406, + 0.021525941789150238, + 0.010441175661981106, + 0.007478472776710987, + 0.006841890048235655, + 0.009229836985468864, + 0.0070198834873735905, + 0.0047194198705255985, + 0.006373629439622164, + 0.011734701693058014, + 0.006913771387189627, + 0.00807231105864048, + 0.0032713001128286123, + 0.005550024099647999, + 0.003836696967482567, + 0.0035954155027866364, + 0.004060703329741955, + 0.010998470708727837, + 0.008930662646889687, + 0.01707654818892479, + 0.012987987138330936, + 0.015577769838273525, + 0.01426154188811779, + 0.00671929307281971, + 0.011426795274019241, + 0.01917978748679161, + 0.010281386785209179, + 0.008019239641726017, + 0.007051693741232157, + 0.005671536084264517, + 0.006182064767926931, + 0.005819108337163925, + 0.006200463976711035, + 0.009810463525354862, + 0.009133903309702873, + 0.04998699203133583, + 0.010027686133980751, + 0.009441960602998734, + 0.010740935802459717, + 0.041364725679159164, + 0.017426300793886185, + 0.01719527505338192, + 0.009736529551446438, + 0.023791184648871422, + 0.01628972217440605, + 0.0176695603877306, + 0.0164725873619318, + 0.025559907779097557, + 0.027403095737099648, + 0.02455185353755951, + 0.031066814437508583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.013290622271597385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.018204864114522934 + ], + "opposite_logprob_tokens": [ + -4.018204689025879 + ], + "logprob_reward": -0.018204864114522934, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018204864114522934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24847593903541565, + 0.08278114348649979, + 0.027551060542464256, + 0.04964037984609604, + 0.023478250950574875, + 0.04141192138195038, + 0.06230509281158447, + 0.043404269963502884, + 0.15128639340400696, + 0.26966553926467896 + ], + "credit_map": [ + 0.04336056113243103, + 0.020331498235464096, + 0.025620050728321075, + 0.022782456129789352, + 0.01765686459839344, + 0.018117714673280716, + 0.05462102219462395, + 0.045985765755176544, + 0.013776296749711037, + 0.009024761617183685, + 0.011268083937466145, + 0.008930793963372707, + 0.0051635028794407845, + 0.007956022396683693, + 0.015483610332012177, + 0.011178082786500454, + 0.003505570348352194, + 0.005437294952571392, + 0.005006503313779831, + 0.003466963302344084, + 0.0018587829545140266, + 0.0017641886370256543, + 0.004205209203064442, + 0.0023065488785505295, + 0.008471569046378136, + 0.00589952664449811, + 0.007978052832186222, + 0.0060515087097883224, + 0.0034397433046251535, + 0.007161041721701622, + 0.004979231394827366, + 0.005659705027937889, + 0.0036003226414322853, + 0.0030892211943864822, + 0.0036234732251614332, + 0.0031859315931797028, + 0.001611949410289526, + 0.0019769337959587574, + 0.0038267667405307293, + 0.002563652815297246, + 0.005866626277565956, + 0.0026341648772358894, + 0.005098178517073393, + 0.0034287087619304657, + 0.0028576150070875883, + 0.003610563464462757, + 0.010232253931462765, + 0.007683807052671909, + 0.008300475776195526, + 0.005440874956548214, + 0.008337356150150299, + 0.006772616878151894, + 0.004537797532975674, + 0.006466716527938843, + 0.0136474734172225, + 0.008801785297691822, + 0.0054925973527133465, + 0.006717849522829056, + 0.005618781317025423, + 0.0061021591536700726, + 0.0029311340767890215, + 0.003084249794483185, + 0.006130799185484648, + 0.007326699793338776, + 0.0501214973628521, + 0.007792165037244558, + 0.008563706651329994, + 0.007350036408752203, + 0.04518576338887215, + 0.01496114768087864, + 0.010126669891178608, + 0.007185404654592276, + 0.02588566765189171, + 0.04220132902264595, + 0.029372284188866615, + 0.026991795748472214, + 0.07116001844406128, + 0.028405355289578438, + 0.027527710422873497, + 0.018121380358934402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.018204864114522934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.009784595109522343 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784595109522343, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784595109522343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16075482964515686, + 0.16820089519023895, + 0.05354958027601242, + 0.04127667844295502, + 0.13786199688911438, + 0.12175343930721283, + 0.06400639563798904, + 0.031028222292661667, + 0.08806407451629639, + 0.13350385427474976 + ], + "credit_map": [ + 0.02005324885249138, + 0.016753604635596275, + 0.022331679239869118, + 0.0169040746986866, + 0.012743894942104816, + 0.016429821029305458, + 0.03787308931350708, + 0.017665421590209007, + 0.027574116364121437, + 0.009610697627067566, + 0.017324300482869148, + 0.012334826402366161, + 0.01166681107133627, + 0.013418827205896378, + 0.045859187841415405, + 0.03041212446987629, + 0.008672860451042652, + 0.0042258151806890965, + 0.005329412408173084, + 0.004007855895906687, + 0.003734615631401539, + 0.005430945660918951, + 0.014781497418880463, + 0.007366575766354799, + 0.007053396664559841, + 0.003970019984990358, + 0.004916470497846603, + 0.004299469757825136, + 0.00271264323964715, + 0.0050122435204684734, + 0.007694338448345661, + 0.00561809865757823, + 0.024373505264520645, + 0.00651645939797163, + 0.015105432830750942, + 0.008408359251916409, + 0.009410114958882332, + 0.009978249669075012, + 0.03345801308751106, + 0.030611861497163773, + 0.019204581156373024, + 0.005923803895711899, + 0.013873090036213398, + 0.00840012263506651, + 0.00867170188575983, + 0.008247816935181618, + 0.03210371732711792, + 0.025328602641820908, + 0.008257092908024788, + 0.007017518859356642, + 0.009414627216756344, + 0.0077825370244681835, + 0.004546404350548983, + 0.00641020480543375, + 0.010304171591997147, + 0.010273834690451622, + 0.005162854678928852, + 0.0037050063256174326, + 0.00337624573148787, + 0.003172385273501277, + 0.0020405903924256563, + 0.0024857802782207727, + 0.0046563162468373775, + 0.006429043598473072, + 0.022847024723887444, + 0.0056656464003026485, + 0.0076292334124445915, + 0.0063184225000441074, + 0.015529142692685127, + 0.008734765462577343, + 0.012636920437216759, + 0.008702920749783516, + 0.014898371882736683, + 0.007428106386214495, + 0.012482323683798313, + 0.010891811922192574, + 0.019102225080132484, + 0.020565509796142578, + 0.020369097590446472, + 0.027766410261392593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.009784595109522343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.16044537723064423 + ], + "opposite_logprob_tokens": [ + -1.9104453325271606 + ], + "logprob_reward": -0.16044537723064423, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16044537723064423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18561629951000214, + 0.14726786315441132, + 0.07330396771430969, + 0.11769595742225647, + 0.01760382205247879, + 0.04725032299757004, + 0.033613041043281555, + 0.026892991736531258, + 0.12944793701171875, + 0.22130781412124634 + ], + "credit_map": [ + 0.03457189351320267, + 0.020192094147205353, + 0.021483812481164932, + 0.01892121694982052, + 0.016552362591028214, + 0.023632369935512543, + 0.026435067877173424, + 0.023827478289604187, + 0.02366514876484871, + 0.010043970309197903, + 0.014269853942096233, + 0.011148645542562008, + 0.009939816780388355, + 0.011202801018953323, + 0.037438709288835526, + 0.02955891191959381, + 0.012827152386307716, + 0.005724627524614334, + 0.008573755621910095, + 0.006165198981761932, + 0.0047784605994820595, + 0.005692543461918831, + 0.01608419045805931, + 0.013458033092319965, + 0.02130468189716339, + 0.00562569173052907, + 0.013442948460578918, + 0.00868312269449234, + 0.008419536054134369, + 0.009582179598510265, + 0.026464786380529404, + 0.024173012003302574, + 0.0024300480727106333, + 0.0023603378795087337, + 0.0021409234032034874, + 0.0015679490752518177, + 0.001414941274560988, + 0.001836551004089415, + 0.0030503852758556604, + 0.0028026863001286983, + 0.007742679212242365, + 0.006370908115059137, + 0.00774429040029645, + 0.0067271944135427475, + 0.003350247861817479, + 0.005974142346531153, + 0.004337203688919544, + 0.005003658588975668, + 0.004244989715516567, + 0.003120959037914872, + 0.004337342921644449, + 0.0032135697547346354, + 0.0024177117738872766, + 0.0038546104915440083, + 0.007751103024929762, + 0.0046727536246180534, + 0.0037457526195794344, + 0.0035656567197293043, + 0.003046790836378932, + 0.003519457997754216, + 0.0021344346459954977, + 0.002715089125558734, + 0.004346148110926151, + 0.003819661447778344, + 0.028793485835194588, + 0.007927490398287773, + 0.010235721245408058, + 0.0094321733340621, + 0.026910174638032913, + 0.01749156042933464, + 0.013078844174742699, + 0.015578490681946278, + 0.02607460878789425, + 0.01781194470822811, + 0.024556977674365044, + 0.01607424207031727, + 0.04608592391014099, + 0.03911612182855606, + 0.024925384670495987, + 0.026662632822990417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.16044537723064423 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18838533759117126, + 0.1046433374285698, + 0.06887348741292953, + 0.04234045743942261, + 0.08544860780239105, + 0.09920281916856766, + 0.03126534819602966, + 0.03417862206697464, + 0.13901998102664948, + 0.2066420018672943 + ], + "credit_map": [ + 0.026752624660730362, + 0.029182258993387222, + 0.028340235352516174, + 0.023281320929527283, + 0.015498313121497631, + 0.017734479159116745, + 0.03374844416975975, + 0.01384763978421688, + 0.018717484548687935, + 0.0075997523963451385, + 0.011210447177290916, + 0.008459296077489853, + 0.006056133657693863, + 0.008659262210130692, + 0.017148256301879883, + 0.026792697608470917, + 0.012833154760301113, + 0.005655590910464525, + 0.009386306628584862, + 0.005963497795164585, + 0.00436099199578166, + 0.007704835385084152, + 0.011582804843783379, + 0.011386299505829811, + 0.006615534890443087, + 0.0030632587149739265, + 0.006333496421575546, + 0.004179741721600294, + 0.002510110381990671, + 0.006325214635580778, + 0.009362252429127693, + 0.003950841259211302, + 0.013142654672265053, + 0.008455081842839718, + 0.010033978149294853, + 0.00686507485806942, + 0.005569141823798418, + 0.005786747206002474, + 0.019613953307271004, + 0.015981966629624367, + 0.013964645564556122, + 0.007832106202840805, + 0.011564177460968494, + 0.008150008507072926, + 0.00695215817540884, + 0.008127560839056969, + 0.025643469765782356, + 0.016968687996268272, + 0.003931886050850153, + 0.0029080784879624844, + 0.003843482118099928, + 0.002611419651657343, + 0.0023138730321079493, + 0.003137477207928896, + 0.00875234417617321, + 0.003766785142943263, + 0.004262762609869242, + 0.004202466923743486, + 0.003706174436956644, + 0.003417334286496043, + 0.0036368491128087044, + 0.004284523893147707, + 0.005989976692944765, + 0.00467853294685483, + 0.03957069665193558, + 0.010137917473912239, + 0.00903038214892149, + 0.009083057753741741, + 0.036734215915203094, + 0.013278360478579998, + 0.01291754748672247, + 0.008267799392342567, + 0.021870093420147896, + 0.01694013737142086, + 0.031409602612257004, + 0.023686377331614494, + 0.02106354385614395, + 0.024470971897244453, + 0.025522606447339058, + 0.041678640991449356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01108264084905386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, focuses on a weary female researcher seated at a cluttered lab table covered with lab equipment. Her dark brown hair is casually tied into a messy bun, a few strands clinging to her cheeks with sweat. A pair of clear goggles are pushed up on her forehead, revealing her sunken eyes and faint dark circles. She wears a slightly wrinkled white lab coat, the collar of which peeks out from under a gray crew-neck T-shirt. Before her, a row of glass beakers and Erlenmeyer flasks hold liquids of various colors—one a vibrant turquoise, another a deep amber—several of which emit a constant, fine bubble and a soft hissing sound. In the background, the cold white light of overhead fluorescent lights fills the room, casting harsh reflections on the stainless steel equipment and glassware. Her shoulders are slightly hunched forward from prolonged concentration, and her gaze, initially blank, stares at the bubbling liquid before her, but her eyelids grow increasingly heavy. Finally, she seemed to give up resistance and slowly, almost reverently, closed her eyes. Her tense jawline finally relaxed at this moment, and she was immersed in a brief but profound tranquility, which contrasted sharply with the ongoing chemical reaction around her, creating an atmosphere of concentration and human vulnerability intertwined between scientific rigor and human exhaustion.", + "seed": 17397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_12/index_6526_seed_17397_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": -0.09717985410243273, + "vlm_logprob": -0.09717985410243273, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.09717985410243273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.35147586464881897 + ], + "opposite_logprob_tokens": [ + -1.2264758348464966 + ], + "logprob_reward": -0.35147586464881897, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35147586464881897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3195822536945343, + 0.10354376584291458, + 0.08966310322284698, + 0.07455462962388992, + 0.0215742364525795, + 0.03019324131309986, + 0.07724256813526154, + 0.07437340915203094, + 0.13363921642303467, + 0.07563353329896927 + ], + "credit_map": [ + 0.03357498720288277, + 0.02830563858151436, + 0.02800520695745945, + 0.05477585643529892, + 0.06245598942041397, + 0.029400324448943138, + 0.04937416687607765, + 0.03369007632136345, + 0.008076501078903675, + 0.013856688514351845, + 0.010205905884504318, + 0.017735745757818222, + 0.015347146429121494, + 0.009884042665362358, + 0.0197459664195776, + 0.00869177095592022, + 0.010225703939795494, + 0.013222597539424896, + 0.009353863075375557, + 0.019830388948321342, + 0.007664068136364222, + 0.008398079313337803, + 0.01333685778081417, + 0.007631547283381224, + 0.0073099141009151936, + 0.009342120960354805, + 0.010372149758040905, + 0.019018100574612617, + 0.006662463769316673, + 0.00738875288516283, + 0.0066812909208238125, + 0.007779841311275959, + 0.00186117400880903, + 0.0029211395885795355, + 0.001875639776699245, + 0.0026394145097583532, + 0.003463525790721178, + 0.0018547257641330361, + 0.004334662109613419, + 0.0026239552535116673, + 0.0018935363041236997, + 0.003957256209105253, + 0.002455382142215967, + 0.0062419939786195755, + 0.004631713032722473, + 0.002779090777039528, + 0.004697761032730341, + 0.0035365084186196327, + 0.014355339109897614, + 0.011819091625511646, + 0.009572828188538551, + 0.011842482723295689, + 0.0038273476529866457, + 0.00800112821161747, + 0.008573469705879688, + 0.009250884875655174, + 0.00505007803440094, + 0.008136125281453133, + 0.010611315257847309, + 0.01028481125831604, + 0.009834043681621552, + 0.007014388684183359, + 0.0135346045717597, + 0.009908045642077923, + 0.007962382398545742, + 0.016372885555028915, + 0.010687636211514473, + 0.028167428448796272, + 0.02920195274055004, + 0.012963168323040009, + 0.016585543751716614, + 0.0116982227191329, + 0.0054014381021261215, + 0.008195615373551846, + 0.011274412274360657, + 0.01709776185452938, + 0.008440651930868626, + 0.00792765710502863, + 0.008845672011375427, + 0.008450325578451157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.35147586464881897 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.06209242716431618 + ], + "opposite_logprob_tokens": [ + -2.8120925426483154 + ], + "logprob_reward": -0.06209242716431618, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209242716431618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34881579875946045, + 0.08244878053665161, + 0.05263855680823326, + 0.09599000960588455, + 0.0330842100083828, + 0.031686365604400635, + 0.05761876702308655, + 0.10112859308719635, + 0.1009855642914772, + 0.09560336917638779 + ], + "credit_map": [ + 0.0559115931391716, + 0.051854897290468216, + 0.03591196984052658, + 0.059965893626213074, + 0.0386236347258091, + 0.03287656977772713, + 0.03641994670033455, + 0.037251267582178116, + 0.011269215494394302, + 0.00986874382942915, + 0.00827725138515234, + 0.012300689704716206, + 0.012003524228930473, + 0.007737356703728437, + 0.01246960461139679, + 0.008522399701178074, + 0.004991831257939339, + 0.005386010278016329, + 0.006088134832680225, + 0.004498074296861887, + 0.007751136552542448, + 0.005058689974248409, + 0.012727904133498669, + 0.00613677641376853, + 0.007745084352791309, + 0.011227822862565517, + 0.010483411140739918, + 0.02800755575299263, + 0.0066998424008488655, + 0.00947363581508398, + 0.013245819136500359, + 0.009106839075684547, + 0.0026145270094275475, + 0.005257575307041407, + 0.0028123543597757816, + 0.005756798665970564, + 0.0047480459325015545, + 0.0033715979661792517, + 0.004581317771226168, + 0.00394199276342988, + 0.00405649421736598, + 0.004362012725323439, + 0.0030098287388682365, + 0.006384978536516428, + 0.0033716631587594748, + 0.0032992379274219275, + 0.003336709225550294, + 0.0038654443342238665, + 0.00898672267794609, + 0.007428388111293316, + 0.0096246637403965, + 0.009681520983576775, + 0.00429791584610939, + 0.005360810086131096, + 0.005745231173932552, + 0.006493512541055679, + 0.007408399134874344, + 0.008673577569425106, + 0.01615159586071968, + 0.027516381815075874, + 0.010378187522292137, + 0.009588852524757385, + 0.012042853981256485, + 0.009368740022182465, + 0.0074905408546328545, + 0.012008544988930225, + 0.016546577215194702, + 0.0177750401198864, + 0.014443489722907543, + 0.00915460754185915, + 0.010506417602300644, + 0.013060351833701134, + 0.00791836902499199, + 0.010111664421856403, + 0.015007797628641129, + 0.014179018326103687, + 0.014219931326806545, + 0.009923296980559826, + 0.013521037064492702, + 0.010722256265580654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.06209242716431618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.008643713779747486 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643713779747486, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643713779747486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23274686932563782, + 0.10919405519962311, + 0.058690160512924194, + 0.03135107830166817, + 0.0773780420422554, + 0.06244669482111931, + 0.15141136944293976, + 0.11256324499845505, + 0.11305808275938034, + 0.051160383969545364 + ], + "credit_map": [ + 0.026678716763854027, + 0.02801239676773548, + 0.03527301922440529, + 0.02296586148440838, + 0.036180052906274796, + 0.02055223286151886, + 0.04290737956762314, + 0.020177215337753296, + 0.00997863058000803, + 0.012337325140833855, + 0.01402400154620409, + 0.019949113950133324, + 0.009938711300492287, + 0.009577019140124321, + 0.019885368645191193, + 0.013503879308700562, + 0.005519510712474585, + 0.005464651621878147, + 0.012859100475907326, + 0.008938072249293327, + 0.004790095146745443, + 0.005277804099023342, + 0.009133887477219105, + 0.006707036402076483, + 0.002516448963433504, + 0.0036535682156682014, + 0.004848991055041552, + 0.005145433358848095, + 0.0031412909738719463, + 0.0032195730600506067, + 0.005527064669877291, + 0.003298706142231822, + 0.0056756408885121346, + 0.00792318768799305, + 0.00941451359540224, + 0.024381378665566444, + 0.0035197169054299593, + 0.008470263332128525, + 0.009860669262707233, + 0.008132676593959332, + 0.007702786009758711, + 0.006878399755805731, + 0.012076410464942455, + 0.014672107994556427, + 0.0038770625833421946, + 0.00601086625829339, + 0.005065415985882282, + 0.006163643207401037, + 0.029297327622771263, + 0.022623680531978607, + 0.020091861486434937, + 0.02772694267332554, + 0.006864236202090979, + 0.015043403953313828, + 0.013097817078232765, + 0.016666095703840256, + 0.009858689270913601, + 0.01021167729049921, + 0.01583622395992279, + 0.029059873893857002, + 0.01013271789997816, + 0.011185775510966778, + 0.012734141200780869, + 0.013544156216084957, + 0.016396421939134598, + 0.019999844953417778, + 0.012348939664661884, + 0.012225531041622162, + 0.019320927560329437, + 0.009909404441714287, + 0.010920044034719467, + 0.011936967261135578, + 0.0065820179879665375, + 0.005451040808111429, + 0.0064269001595675945, + 0.005380298942327499, + 0.008392839692533016, + 0.004756927024573088, + 0.009158959612250328, + 0.005011402070522308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.008643713779747486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.04309491440653801 + ], + "opposite_logprob_tokens": [ + -3.1680948734283447 + ], + "logprob_reward": -0.04309491440653801, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309491440653801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04542534798383713, + 0.05783341825008392, + 0.020650912076234818, + 0.01039703655987978, + 0.04238024726510048, + 0.017686480656266212, + 0.07352790981531143, + 0.1292460560798645, + 0.4761011302471161, + 0.12675146758556366 + ], + "credit_map": [ + 0.0043212538585066795, + 0.005742980632930994, + 0.007635527290403843, + 0.0056444210931658745, + 0.004629964008927345, + 0.00531674874946475, + 0.007407004479318857, + 0.0047274441458284855, + 0.004824075382202864, + 0.005388897377997637, + 0.00856000930070877, + 0.01574450358748436, + 0.0034020289313048124, + 0.006245758384466171, + 0.00779731897637248, + 0.005870814900845289, + 0.002763493452221155, + 0.005086096003651619, + 0.002386606065556407, + 0.002601456129923463, + 0.0015109189553186297, + 0.0024576461873948574, + 0.002026718109846115, + 0.0018179757753387094, + 0.0007190757314674556, + 0.0011002301471307874, + 0.0013351290253922343, + 0.002532795537263155, + 0.0007538000354543328, + 0.0010519165080040693, + 0.0015211396384984255, + 0.001382949179969728, + 0.0032794757280498743, + 0.0035731710959225893, + 0.006410188507288694, + 0.012550308369100094, + 0.002017923165112734, + 0.0045110490173101425, + 0.00565356994047761, + 0.0043845572508871555, + 0.0013369316002354026, + 0.0020477029029279947, + 0.0020009016152471304, + 0.00505713140591979, + 0.001263661659322679, + 0.002210096688941121, + 0.001990159507840872, + 0.0017798937624320388, + 0.013760586269199848, + 0.011199835687875748, + 0.009112277068197727, + 0.012794461101293564, + 0.003381668124347925, + 0.007384751457720995, + 0.00769227696582675, + 0.008202046155929565, + 0.008805413730442524, + 0.010175563395023346, + 0.02284833788871765, + 0.03731381520628929, + 0.008792348206043243, + 0.012235704809427261, + 0.015549938194453716, + 0.013524921610951424, + 0.07979287952184677, + 0.0910930335521698, + 0.04604300111532211, + 0.07876784354448318, + 0.04746513441205025, + 0.03940175846219063, + 0.0404670424759388, + 0.05307040363550186, + 0.02627832069993019, + 0.01237522903829813, + 0.02240769937634468, + 0.013134518638253212, + 0.012713141739368439, + 0.015442384406924248, + 0.011847171001136303, + 0.012552989646792412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.04309491440653801 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.020592350512742996 + ], + "opposite_logprob_tokens": [ + -3.895592451095581 + ], + "logprob_reward": -0.020592350512742996, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592350512742996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26607292890548706, + 0.0890195369720459, + 0.04315720871090889, + 0.03878955543041229, + 0.02454763650894165, + 0.03897685930132866, + 0.04952721670269966, + 0.05469164624810219, + 0.14505204558372498, + 0.2501654326915741 + ], + "credit_map": [ + 0.03500979393720627, + 0.031478751450777054, + 0.026125626638531685, + 0.020399007946252823, + 0.052031662315130234, + 0.027506299316883087, + 0.040684424340724945, + 0.03283735737204552, + 0.012406667694449425, + 0.012274714186787605, + 0.010091934353113174, + 0.010415304452180862, + 0.013682153075933456, + 0.008225596509873867, + 0.01240555103868246, + 0.009517614729702473, + 0.0050597726367414, + 0.007278065197169781, + 0.004575179424136877, + 0.005041930358856916, + 0.0064215888269245625, + 0.004324097651988268, + 0.0058602215722203255, + 0.004596351180225611, + 0.004917428363114595, + 0.004919604863971472, + 0.004486523102968931, + 0.005083765834569931, + 0.003646192839369178, + 0.004133492708206177, + 0.0067827701568603516, + 0.004819776862859726, + 0.0029717618599534035, + 0.0030175005085766315, + 0.002752334577962756, + 0.004940155427902937, + 0.0026357686147093773, + 0.0024697785265743732, + 0.002804493997246027, + 0.0029558443930000067, + 0.0036200929898768663, + 0.0037105625960975885, + 0.0045347814448177814, + 0.010055405087769032, + 0.0033080822322517633, + 0.003978280816227198, + 0.005325862672179937, + 0.004443792626261711, + 0.004509808961302042, + 0.006473550107330084, + 0.005973610561341047, + 0.00826446246355772, + 0.006645566318184137, + 0.005343640688806772, + 0.005269581917673349, + 0.007046996615827084, + 0.004327038303017616, + 0.006832093931734562, + 0.006161245983093977, + 0.012003758922219276, + 0.006629989016801119, + 0.0048813908360898495, + 0.006837166845798492, + 0.007018962875008583, + 0.0190459992736578, + 0.023876965045928955, + 0.022093432024121284, + 0.017208727076649666, + 0.02148784138262272, + 0.012539872899651527, + 0.015553778037428856, + 0.013245435431599617, + 0.021775037050247192, + 0.02803276665508747, + 0.033156681805849075, + 0.06047653406858444, + 0.029231319203972816, + 0.022507304325699806, + 0.03167394921183586, + 0.02331184409558773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.020592350512742996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, focuses on a weary female researcher seated at a cluttered lab table covered with lab equipment. Her dark brown hair is casually tied into a messy bun, a few strands clinging to her cheeks with sweat. A pair of clear goggles are pushed up on her forehead, revealing her sunken eyes and faint dark circles. She wears a slightly wrinkled white lab coat, the collar of which peeks out from under a gray crew-neck T-shirt. Before her, a row of glass beakers and Erlenmeyer flasks hold liquids of various colors—one a vibrant turquoise, another a deep amber—several of which emit a constant, fine bubble and a soft hissing sound. In the background, the cold white light of overhead fluorescent lights fills the room, casting harsh reflections on the stainless steel equipment and glassware. Her shoulders are slightly hunched forward from prolonged concentration, and her gaze, initially blank, stares at the bubbling liquid before her, but her eyelids grow increasingly heavy. Finally, she seemed to give up resistance and slowly, almost reverently, closed her eyes. Her tense jawline finally relaxed at this moment, and she was immersed in a brief but profound tranquility, which contrasted sharply with the ongoing chemical reaction around her, creating an atmosphere of concentration and human vulnerability intertwined between scientific rigor and human exhaustion.", + "seed": 17398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_12/index_6526_seed_17398_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": -0.025321345124393702, + "vlm_logprob": -0.025321345124393702, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.025321345124393702, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.006018373183906078 + ], + "opposite_logprob_tokens": [ + -5.756018161773682 + ], + "logprob_reward": -0.006018373183906078, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.006018373183906078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11412682384252548, + 0.0766381323337555, + 0.05013253539800644, + 0.06118840351700783, + 0.06183907017111778, + 0.1476704627275467, + 0.1938689649105072, + 0.1653173416852951, + 0.038296572864055634, + 0.09092167764902115 + ], + "credit_map": [ + 0.02055473066866398, + 0.013195446692407131, + 0.01151934266090393, + 0.013270397670567036, + 0.01491357758641243, + 0.01678212732076645, + 0.0116084273904562, + 0.012282773852348328, + 0.009568634442985058, + 0.009512710385024548, + 0.007266297470778227, + 0.009395617991685867, + 0.01425238512456417, + 0.008235163986682892, + 0.013442452065646648, + 0.004964870400726795, + 0.005827278830111027, + 0.005874180234968662, + 0.004789788741618395, + 0.0060475291684269905, + 0.007202540058642626, + 0.007391269784420729, + 0.007444493938237429, + 0.0055554574355483055, + 0.011378082446753979, + 0.0067152283154428005, + 0.004539544694125652, + 0.005622093100100756, + 0.011110613122582436, + 0.010997256264090538, + 0.007373622618615627, + 0.003451964119449258, + 0.007671163883060217, + 0.006992971058934927, + 0.008422577753663063, + 0.00872600544244051, + 0.007519923150539398, + 0.010997827164828777, + 0.005545689258724451, + 0.005962916184216738, + 0.019235199317336082, + 0.0155472531914711, + 0.01812080293893814, + 0.01907246746122837, + 0.023901615291833878, + 0.0207088403403759, + 0.016191624104976654, + 0.014892646111547947, + 0.024921325966715813, + 0.027922671288251877, + 0.017011502757668495, + 0.01717345044016838, + 0.03664834052324295, + 0.02676648646593094, + 0.023784667253494263, + 0.01964053139090538, + 0.022127944976091385, + 0.021260317414999008, + 0.018379082903265953, + 0.018217772245407104, + 0.020678363740444183, + 0.033224690705537796, + 0.014706958085298538, + 0.016722215339541435, + 0.004791338462382555, + 0.004348792135715485, + 0.0035597169771790504, + 0.0058722286485135555, + 0.0068668401800096035, + 0.004280885681509972, + 0.004943923559039831, + 0.0036328460555523634, + 0.011113930493593216, + 0.012507935985922813, + 0.012030043639242649, + 0.011155357584357262, + 0.014440029859542847, + 0.011061746627092361, + 0.009622994810342789, + 0.00898964237421751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.006018373183906078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.014200118370354176 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200118370354176, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200118370354176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24947573244571686, + 0.09512919187545776, + 0.051172930747270584, + 0.04456734657287598, + 0.05589917674660683, + 0.07647992670536041, + 0.05920848622918129, + 0.09684769809246063, + 0.10571909695863724, + 0.1655004322528839 + ], + "credit_map": [ + 0.04858534410595894, + 0.0250454880297184, + 0.020672427490353584, + 0.021151507273316383, + 0.029750360175967216, + 0.049385473132133484, + 0.02979222871363163, + 0.025092896074056625, + 0.016515417024493217, + 0.012621414847671986, + 0.010892120189964771, + 0.010203340090811253, + 0.010748215951025486, + 0.013394410721957684, + 0.01046890951693058, + 0.01028535608202219, + 0.007559731602668762, + 0.005818567704409361, + 0.005263044964522123, + 0.0057735745795071125, + 0.008533559739589691, + 0.005795436445623636, + 0.006318447180092335, + 0.006110568530857563, + 0.008278980851173401, + 0.004623146262019873, + 0.0035980052780359983, + 0.004489589016884565, + 0.006068250630050898, + 0.006150089669972658, + 0.007274553179740906, + 0.004084731452167034, + 0.008727055042982101, + 0.005560750141739845, + 0.006224796641618013, + 0.008381842635571957, + 0.006561845541000366, + 0.009389864280819893, + 0.0047601391561329365, + 0.0062928833067417145, + 0.010390643030405045, + 0.007012808695435524, + 0.00859057530760765, + 0.008770917542278767, + 0.008486740291118622, + 0.012218929827213287, + 0.008182616904377937, + 0.012826696038246155, + 0.008344843983650208, + 0.0074490513652563095, + 0.004611099138855934, + 0.007194768637418747, + 0.011666853912174702, + 0.0063717542216181755, + 0.007947007194161415, + 0.0056231035850942135, + 0.013961640186607838, + 0.011412340216338634, + 0.009229421615600586, + 0.013569897040724754, + 0.01399620994925499, + 0.015605662018060684, + 0.010986561886966228, + 0.008085963316261768, + 0.013540373183786869, + 0.01727764680981636, + 0.011095681227743626, + 0.013285455293953419, + 0.01231511402875185, + 0.010966016910970211, + 0.01924915798008442, + 0.007989655248820782, + 0.013152000494301319, + 0.02032872475683689, + 0.02029496245086193, + 0.024199357256293297, + 0.02390686236321926, + 0.017860213294625282, + 0.027508556842803955, + 0.018249748274683952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.014200118370354176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.004096450749784708 + ], + "opposite_logprob_tokens": [ + -5.504096508026123 + ], + "logprob_reward": -0.004096450749784708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096450749784708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26448458433151245, + 0.11998332291841507, + 0.12039569765329361, + 0.04053264111280441, + 0.08638785034418106, + 0.11312218010425568, + 0.0698462426662445, + 0.05794361233711243, + 0.07556162029504776, + 0.05174218490719795 + ], + "credit_map": [ + 0.041125841438770294, + 0.030334535986185074, + 0.031839776784181595, + 0.030524762347340584, + 0.026081906631588936, + 0.04746513068675995, + 0.02854011580348015, + 0.028572550043463707, + 0.015542627312242985, + 0.013383131474256516, + 0.012341893278062344, + 0.02046862617135048, + 0.011991010047495365, + 0.018234824761748314, + 0.013979380019009113, + 0.014041840098798275, + 0.021508172154426575, + 0.012090166099369526, + 0.009326301515102386, + 0.018567312508821487, + 0.014885988086462021, + 0.023335495963692665, + 0.009678077884018421, + 0.01100418996065855, + 0.004842629190534353, + 0.004805115517228842, + 0.005090177524834871, + 0.0082479203119874, + 0.004524999763816595, + 0.0058516934514045715, + 0.0033134028781205416, + 0.003856706665828824, + 0.011076002381742, + 0.009765803813934326, + 0.01050060149282217, + 0.010577867738902569, + 0.012835072353482246, + 0.01418822631239891, + 0.009477928280830383, + 0.007966352626681328, + 0.01592944748699665, + 0.01132735051214695, + 0.013527371920645237, + 0.018608998507261276, + 0.008816011250019073, + 0.016566608101129532, + 0.01167614758014679, + 0.01667024753987789, + 0.011730236932635307, + 0.00685025192797184, + 0.009456999599933624, + 0.010687388479709625, + 0.007286252919584513, + 0.009162914007902145, + 0.005466633941978216, + 0.009205569513142109, + 0.007605114486068487, + 0.0064661214128136635, + 0.004476155154407024, + 0.009449813514947891, + 0.010222875513136387, + 0.0065852804109454155, + 0.006676838267594576, + 0.006461413111537695, + 0.011063167825341225, + 0.013368584215641022, + 0.00686056911945343, + 0.007266896776854992, + 0.013639433309435844, + 0.007624219171702862, + 0.010396572761237621, + 0.005342183634638786, + 0.006883072666823864, + 0.006376714911311865, + 0.005340214353054762, + 0.007438894361257553, + 0.008020443841814995, + 0.006579369306564331, + 0.006646723486483097, + 0.004456756170839071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004096450749784708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.07900068163871765 + ], + "opposite_logprob_tokens": [ + -2.57900071144104 + ], + "logprob_reward": -0.07900068163871765, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900068163871765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21783499419689178, + 0.08113585412502289, + 0.06278786808252335, + 0.06103202700614929, + 0.042987924069166183, + 0.16154102981090546, + 0.17964525520801544, + 0.113707534968853, + 0.03298386558890343, + 0.046343639492988586 + ], + "credit_map": [ + 0.04042132943868637, + 0.023054851219058037, + 0.01906767673790455, + 0.016582751646637917, + 0.034872520714998245, + 0.043294958770275116, + 0.024235712364315987, + 0.01630519889295101, + 0.008980675600469112, + 0.008477280847728252, + 0.006077656522393227, + 0.013127676211297512, + 0.010164985433220863, + 0.012446380220353603, + 0.011657985858619213, + 0.010203211568295956, + 0.0120145408436656, + 0.005706372205168009, + 0.00398248853161931, + 0.010862849652767181, + 0.005246546119451523, + 0.014813701622188091, + 0.004230957943946123, + 0.005930413957685232, + 0.00796181708574295, + 0.007116017397493124, + 0.005641578696668148, + 0.007523317821323872, + 0.00984119437634945, + 0.007533978670835495, + 0.007754873018711805, + 0.007659249007701874, + 0.006663572043180466, + 0.005040578544139862, + 0.0050309570506215096, + 0.004548836499452591, + 0.005473922472447157, + 0.004820976406335831, + 0.004919730126857758, + 0.006489354185760021, + 0.022302702069282532, + 0.01603056490421295, + 0.02534930221736431, + 0.02339354157447815, + 0.017757927998900414, + 0.018905101343989372, + 0.0152702322229743, + 0.022531658411026, + 0.025082677602767944, + 0.02435622178018093, + 0.015054481104016304, + 0.017704561352729797, + 0.035467881709337234, + 0.021940289065241814, + 0.023241136223077774, + 0.01679801754653454, + 0.015742013230919838, + 0.014008322730660439, + 0.013324630446732044, + 0.01080196164548397, + 0.014209500513970852, + 0.021343879401683807, + 0.010821214877068996, + 0.013456016778945923, + 0.004479586612433195, + 0.005970288533717394, + 0.003935650922358036, + 0.003161150962114334, + 0.004407274071127176, + 0.004193425178527832, + 0.003697912208735943, + 0.003138575702905655, + 0.006412893533706665, + 0.006351755931973457, + 0.006502693984657526, + 0.006446463987231255, + 0.005553331226110458, + 0.005071043502539396, + 0.004617955069988966, + 0.005387500394135714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.07900068163871765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.023291101679205894 + ], + "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": [ + 0.20069248974323273, + 0.1122879609465599, + 0.09448191523551941, + 0.06219746544957161, + 0.042776644229888916, + 0.11729322373867035, + 0.10057667642831802, + 0.06410931795835495, + 0.1252698451280594, + 0.08031445741653442 + ], + "credit_map": [ + 0.022715725004673004, + 0.022586295381188393, + 0.021304184570908546, + 0.026301367208361626, + 0.045646872371435165, + 0.02273198403418064, + 0.028663817793130875, + 0.010742250829935074, + 0.019528349861502647, + 0.01280280202627182, + 0.011269028298556805, + 0.013270680792629719, + 0.014590691775083542, + 0.018101533874869347, + 0.014548122882843018, + 0.008176750503480434, + 0.014150435104966164, + 0.013950490392744541, + 0.010896099731326103, + 0.011566643603146076, + 0.010283718816936016, + 0.014582001604139805, + 0.009520146064460278, + 0.009532386437058449, + 0.008316473104059696, + 0.007842390798032284, + 0.004419263917952776, + 0.006573955994099379, + 0.012079823762178421, + 0.009447469376027584, + 0.008139963261783123, + 0.005378120578825474, + 0.005980825517326593, + 0.00507983285933733, + 0.005085863173007965, + 0.005505196284502745, + 0.005343687254935503, + 0.007153560873121023, + 0.0039123124442994595, + 0.004715370014309883, + 0.0147412633523345, + 0.011317193508148193, + 0.017384110018610954, + 0.014728541485965252, + 0.01655883900821209, + 0.016816487535834312, + 0.012257195077836514, + 0.013489601202309132, + 0.014510499313473701, + 0.013641351833939552, + 0.009400684386491776, + 0.008698890917003155, + 0.01843792386353016, + 0.013814507983624935, + 0.011901183053851128, + 0.010171630419790745, + 0.00933386292308569, + 0.00785616971552372, + 0.0071370285004377365, + 0.0061639780178666115, + 0.008398792706429958, + 0.011124537326395512, + 0.00642944872379303, + 0.0076654995791614056, + 0.01392289623618126, + 0.01585400104522705, + 0.009135575965046883, + 0.018059847876429558, + 0.021954672411084175, + 0.015094609931111336, + 0.01877482421696186, + 0.01247341837733984, + 0.008733381517231464, + 0.00979912094771862, + 0.010409967973828316, + 0.010549108497798443, + 0.01381183136254549, + 0.009397507645189762, + 0.008682779036462307, + 0.008930754847824574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.023291101679205894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a level angle, focuses on a weary female researcher seated at a cluttered lab table covered with lab equipment. Her dark brown hair is casually tied into a messy bun, a few strands clinging to her cheeks with sweat. A pair of clear goggles are pushed up on her forehead, revealing her sunken eyes and faint dark circles. She wears a slightly wrinkled white lab coat, the collar of which peeks out from under a gray crew-neck T-shirt. Before her, a row of glass beakers and Erlenmeyer flasks hold liquids of various colors—one a vibrant turquoise, another a deep amber—several of which emit a constant, fine bubble and a soft hissing sound. In the background, the cold white light of overhead fluorescent lights fills the room, casting harsh reflections on the stainless steel equipment and glassware. Her shoulders are slightly hunched forward from prolonged concentration, and her gaze, initially blank, stares at the bubbling liquid before her, but her eyelids grow increasingly heavy. Finally, she seemed to give up resistance and slowly, almost reverently, closed her eyes. Her tense jawline finally relaxed at this moment, and she was immersed in a brief but profound tranquility, which contrasted sharply with the ongoing chemical reaction around her, creating an atmosphere of concentration and human vulnerability intertwined between scientific rigor and human exhaustion.", + "seed": 17399, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_12/index_6526_seed_17399_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": -0.03495260048657656, + "vlm_logprob": -0.03495260048657656, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.03495260048657656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.013894145376980305 + ], + "opposite_logprob_tokens": [ + -4.638894081115723 + ], + "logprob_reward": -0.013894145376980305, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013894145376980305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33448687195777893, + 0.08972859382629395, + 0.04789377748966217, + 0.0510711595416069, + 0.041650477796792984, + 0.02504665218293667, + 0.02438824623823166, + 0.04411613941192627, + 0.0616292804479599, + 0.27998876571655273 + ], + "credit_map": [ + 0.06805039942264557, + 0.024906009435653687, + 0.04783899337053299, + 0.03975899517536163, + 0.03715623915195465, + 0.027465565130114555, + 0.046925753355026245, + 0.042384929955005646, + 0.015858154743909836, + 0.01219005323946476, + 0.009714961051940918, + 0.010330306366086006, + 0.009853310883045197, + 0.009206879884004593, + 0.00985304918140173, + 0.01272187102586031, + 0.009180570021271706, + 0.004857887979596853, + 0.005612672306597233, + 0.0050706216134130955, + 0.006220591254532337, + 0.005545471329241991, + 0.005821916740387678, + 0.005584046710282564, + 0.008296117186546326, + 0.009148916229605675, + 0.005385273601859808, + 0.005489525385200977, + 0.005950696766376495, + 0.00547452550381422, + 0.005257471464574337, + 0.006068635266274214, + 0.007546590641140938, + 0.004186361562460661, + 0.0051432447507977486, + 0.004629136994481087, + 0.005773339420557022, + 0.004175371956080198, + 0.005122753791511059, + 0.005073681939393282, + 0.0034048566594719887, + 0.002640865044668317, + 0.0034045602660626173, + 0.004283490125089884, + 0.003096685977652669, + 0.002571450313553214, + 0.003033718792721629, + 0.0026110257022082806, + 0.003570893546566367, + 0.0035803073551505804, + 0.0030785780400037766, + 0.0031422581523656845, + 0.0030332584865391254, + 0.0027359682135283947, + 0.0020982827991247177, + 0.0031486975494772196, + 0.008795798756182194, + 0.004148414824157953, + 0.006509598810225725, + 0.004013457801192999, + 0.005237494595348835, + 0.0041311271488666534, + 0.005302934441715479, + 0.005977311637252569, + 0.008127475157380104, + 0.005230064038187265, + 0.01032990962266922, + 0.00993745680898428, + 0.007960356771945953, + 0.0056345476768910885, + 0.005959684960544109, + 0.008449784480035305, + 0.042045824229717255, + 0.02853688970208168, + 0.04710448905825615, + 0.039999719709157944, + 0.021735932677984238, + 0.019841911271214485, + 0.037521325051784515, + 0.043202679604291916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.013894145376980305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.03371366113424301 + ], + "opposite_logprob_tokens": [ + -3.4087135791778564 + ], + "logprob_reward": -0.03371366113424301, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371366113424301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32059958577156067, + 0.11026251316070557, + 0.06466097384691238, + 0.0318610817193985, + 0.10919152200222015, + 0.02307533659040928, + 0.0329115092754364, + 0.03169909492135048, + 0.16018414497375488, + 0.11555425077676773 + ], + "credit_map": [ + 0.037229761481285095, + 0.06861566007137299, + 0.029207849875092506, + 0.041379544883966446, + 0.03778647631406784, + 0.036015912890434265, + 0.03664408251643181, + 0.03372030332684517, + 0.01600208505988121, + 0.025272367522120476, + 0.008457629010081291, + 0.014526941813528538, + 0.010212557390332222, + 0.012486615218222141, + 0.009747947566211224, + 0.013556364923715591, + 0.010606586001813412, + 0.008202060125768185, + 0.0058796885423362255, + 0.006928991060703993, + 0.009805425070226192, + 0.0067539154551923275, + 0.008657402358949184, + 0.007826902903616428, + 0.004201819188892841, + 0.0029354530852288008, + 0.003202054649591446, + 0.004013014025986195, + 0.005241082515567541, + 0.0031480230391025543, + 0.005371479783207178, + 0.003748158924281597, + 0.017055019736289978, + 0.016253070905804634, + 0.009872082620859146, + 0.01098998449742794, + 0.016740094870328903, + 0.010415640659630299, + 0.013469365425407887, + 0.014396263286471367, + 0.0032035475596785545, + 0.0018039909191429615, + 0.0026495554484426975, + 0.0031475259456783533, + 0.0038803573697805405, + 0.0016332768136635423, + 0.0038276214618235826, + 0.002929459558799863, + 0.005648870952427387, + 0.003103925148025155, + 0.004760861862450838, + 0.0039091529324650764, + 0.004410679452121258, + 0.0024455843959003687, + 0.004813813138753176, + 0.0038186199963092804, + 0.004652827046811581, + 0.0035363011993467808, + 0.003543276572600007, + 0.005038861185312271, + 0.003781392704695463, + 0.0034011092502623796, + 0.004314290825277567, + 0.0034310377668589354, + 0.012659185566008091, + 0.012653169222176075, + 0.015472695231437683, + 0.024848567321896553, + 0.03204898536205292, + 0.015810232609510422, + 0.029649939388036728, + 0.017041372135281563, + 0.00909570325165987, + 0.007756280712783337, + 0.01369326002895832, + 0.01590447500348091, + 0.019432656466960907, + 0.012900345958769321, + 0.025803346186876297, + 0.010968184098601341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03371366113424301 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.008638394996523857 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638394996523857, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638394996523857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27354133129119873, + 0.08915093541145325, + 0.06823478639125824, + 0.07983219623565674, + 0.13536137342453003, + 0.03875119239091873, + 0.03926748037338257, + 0.029296744614839554, + 0.03664388135075569, + 0.2099199891090393 + ], + "credit_map": [ + 0.06975772976875305, + 0.047359999269247055, + 0.03087485209107399, + 0.025696948170661926, + 0.02318054996430874, + 0.01933532953262329, + 0.016954107210040092, + 0.040381815284490585, + 0.01802651956677437, + 0.011823688633739948, + 0.012214530259370804, + 0.009153144434094429, + 0.006666087545454502, + 0.010229755192995071, + 0.007958474569022655, + 0.013078740797936916, + 0.01380870770663023, + 0.007213745266199112, + 0.009339007548987865, + 0.008464195765554905, + 0.005736862774938345, + 0.01002893503755331, + 0.005359267350286245, + 0.008284065872430801, + 0.018224578350782394, + 0.008333269506692886, + 0.011330773122608662, + 0.009264443069696426, + 0.0051596928387880325, + 0.011887611821293831, + 0.005682248156517744, + 0.009949573315680027, + 0.03448821231722832, + 0.015421214513480663, + 0.018150728195905685, + 0.017684346064925194, + 0.008823358453810215, + 0.013722061179578304, + 0.00847608968615532, + 0.018595362082123756, + 0.005333463195711374, + 0.0037137356121093035, + 0.0060096918605268, + 0.007615974638611078, + 0.0035450654104351997, + 0.003853473113849759, + 0.004068639129400253, + 0.004611148498952389, + 0.006713747512549162, + 0.004497881047427654, + 0.004814465995877981, + 0.003802953287959099, + 0.004754477646201849, + 0.0051263743080198765, + 0.004712669178843498, + 0.004844913259148598, + 0.00685553252696991, + 0.0033323364332318306, + 0.004170164465904236, + 0.0035182801075279713, + 0.002172226319089532, + 0.003101749112829566, + 0.0022120424546301365, + 0.0039344108663499355, + 0.0040197717025876045, + 0.003397757885977626, + 0.006329904310405254, + 0.0061540487222373486, + 0.0041287620551884174, + 0.0035309884697198868, + 0.0045333439484238625, + 0.004549304489046335, + 0.02750692330300808, + 0.017436562106013298, + 0.030525023117661476, + 0.025000547990202904, + 0.02451319247484207, + 0.019152194261550903, + 0.03505993261933327, + 0.03072563000023365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.008638394996523857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.10032832622528076 + ], + "opposite_logprob_tokens": [ + -2.3503284454345703 + ], + "logprob_reward": -0.10032832622528076, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032832622528076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.270283579826355, + 0.06399443000555038, + 0.2754542827606201, + 0.12903839349746704, + 0.019688183441758156, + 0.07678903639316559, + 0.027669265866279602, + 0.05323990806937218, + 0.0225960873067379, + 0.061246804893016815 + ], + "credit_map": [ + 0.06416969746351242, + 0.019265687093138695, + 0.039909008890390396, + 0.032578203827142715, + 0.01663120649755001, + 0.034403976052999496, + 0.018117152154445648, + 0.0452086441218853, + 0.01600460894405842, + 0.008640703745186329, + 0.0073946635238826275, + 0.006490189116448164, + 0.006715388037264347, + 0.005584779195487499, + 0.004726644139736891, + 0.0084374463185668, + 0.04303580895066261, + 0.03130347281694412, + 0.035426750779151917, + 0.0303437989205122, + 0.024494171142578125, + 0.05397659167647362, + 0.02115004137158394, + 0.035723648965358734, + 0.026158183813095093, + 0.024279525503516197, + 0.012872297316789627, + 0.01103924959897995, + 0.011280125007033348, + 0.017878428101539612, + 0.008316870778799057, + 0.017213715240359306, + 0.004317076411098242, + 0.00181498599704355, + 0.0025365017354488373, + 0.0026322375051677227, + 0.001826980384066701, + 0.0022539538331329823, + 0.0017875551711767912, + 0.0025188929866999388, + 0.0159830991178751, + 0.01132403314113617, + 0.008223537355661392, + 0.008408448658883572, + 0.005776256322860718, + 0.011463082395493984, + 0.005614630877971649, + 0.009995946660637856, + 0.003471655771136284, + 0.0051225763745605946, + 0.002431104891002178, + 0.0032473893370479345, + 0.0033754820469766855, + 0.003092811442911625, + 0.0036756987683475018, + 0.003252545138821006, + 0.009181694127619267, + 0.005549257155507803, + 0.007282276637852192, + 0.005812105257064104, + 0.004766594152897596, + 0.00969452690333128, + 0.004222653340548277, + 0.006730802357196808, + 0.002693678718060255, + 0.0024923442397266626, + 0.0028801187872886658, + 0.004206240642815828, + 0.0023316978476941586, + 0.002337019657716155, + 0.0026676729321479797, + 0.0029873144812881947, + 0.006694903131574392, + 0.007710393518209457, + 0.004290263168513775, + 0.006039685569703579, + 0.012114709243178368, + 0.00616172282025218, + 0.010255850851535797, + 0.007979274727404118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.10032832622528076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 close-up of a female researcher seated at a cluttered lab table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the researcher have dark brown hair tied in a messy bun with clear goggles pushed up on her forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there glass beakers and Erleneyer flasks containing liquids of various colors, including vibrant turquoise and deep amber, emitting bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background illuminated by cold white overhead fluorescent lights casting harsh reflections on stainless steel equipment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level 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": [ + -0.01818847469985485 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818847469985485, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818847469985485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26408034563064575, + 0.05236073583364487, + 0.04272511601448059, + 0.044623907655477524, + 0.04701250046491623, + 0.024790680035948753, + 0.10466266423463821, + 0.043354835361242294, + 0.10066623985767365, + 0.27572301030158997 + ], + "credit_map": [ + 0.06868463009595871, + 0.024869535118341446, + 0.037873342633247375, + 0.03044021874666214, + 0.015795383602380753, + 0.02628273330628872, + 0.024827271699905396, + 0.035307224839925766, + 0.0075220572762191296, + 0.0066888644360005856, + 0.007774531841278076, + 0.005595214199274778, + 0.00635804608464241, + 0.005897136405110359, + 0.006758504547178745, + 0.005766380578279495, + 0.005873412359505892, + 0.007485825568437576, + 0.004746655933558941, + 0.0065513732843101025, + 0.004355224315077066, + 0.005199969746172428, + 0.004111023154109716, + 0.004401632584631443, + 0.006823176052421331, + 0.011082288809120655, + 0.0034328100737184286, + 0.003922870382666588, + 0.004488583188503981, + 0.005854024086147547, + 0.0036421024706214666, + 0.005378052592277527, + 0.007457434665411711, + 0.008274503983557224, + 0.005079778376966715, + 0.005185727961361408, + 0.005373283755034208, + 0.004438584204763174, + 0.005918998271226883, + 0.005284191574901342, + 0.0033113318495452404, + 0.005658729467540979, + 0.002281865570694208, + 0.0026809226255863905, + 0.002873000456020236, + 0.0022177889477461576, + 0.002901351312175393, + 0.0028656881768256426, + 0.030589183792471886, + 0.011258825659751892, + 0.018159784376621246, + 0.008035892620682716, + 0.004691284149885178, + 0.010192927904427052, + 0.00556761072948575, + 0.016167154535651207, + 0.004960379563271999, + 0.0042651016265153885, + 0.004769098944962025, + 0.005042977165430784, + 0.006843012757599354, + 0.005402569659054279, + 0.0070639667101204395, + 0.005007729399949312, + 0.014217832125723362, + 0.009394189342856407, + 0.015111775137484074, + 0.01378974225372076, + 0.014645985327661037, + 0.008484712801873684, + 0.013071314431726933, + 0.011950690299272537, + 0.04148655757308006, + 0.02238517627120018, + 0.03802035376429558, + 0.03597448766231537, + 0.04922666400671005, + 0.019660815596580505, + 0.044455114752054214, + 0.02451382763683796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.01818847469985485 + } + ], + "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/0000131/group_13/index_tensor([1204, 1204])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_tensor([1204, 1204])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..eb126d3356fb19c583759742a0e77fc8a25bfdc2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_tensor([1204, 1204])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic low-angle wide shot opens the scene, and the camera begins a slow, deliberate upward tilt, following the momentous ascent of a lone, sleek Martian space elevator pod. As it lifts off, its powerful repulsorgravity engines emit a soft, pulsating blue glow from the vehicle's chassis, silently lifting a vast, swirling cloud of fine ochre dust that billows across the desolate landscape. The pod itself is an aerodynamic teardrop of polished dark gray composite material, its surface unmarked and reflecting the faint sunlight with a subtle sheen. It glides inexorably upward along a vast, shimmering ribbon of semi-translucent material that stretches from the ground to the pale, empty, butterscotch-colored sky, disappearing into the thin upper atmosphere. As the pod climbs higher, its incredibly long, sharp shadow casts a shadow across the cratered terrain below, bisecting the reddish-brown landscape and sweeping across it like the hands of a giant clock. Heavily desaturated color gradients are used throughout key scenes, fading vibrant Martian tones to dusty browns, pale oranges, and cool grays, creating a solemn and awe-inspiring atmosphere that carries profound isolation and historical significance.", + "seed": 29054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_1204_seed_29054_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": -0.062119879946112636, + "vlm_logprob": -0.062119879946112636, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.062119879946112636, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13007056713104248 + ], + "opposite_logprob_tokens": [ + -2.130070686340332 + ], + "logprob_reward": -0.13007056713104248, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13007056713104248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12918134033679962, + 0.1433211714029312, + 0.07730062305927277, + 0.09258068352937698, + 0.261540025472641, + 0.05147811025381088, + 0.08072768151760101, + 0.08229798823595047, + 0.05177900567650795, + 0.02979341521859169 + ], + "credit_map": [ + 0.02117735892534256, + 0.012672965414822102, + 0.015446736477315426, + 0.014789454638957977, + 0.011209189891815186, + 0.010672660544514656, + 0.022853484377264977, + 0.020359482616186142, + 0.02265034057199955, + 0.028157269582152367, + 0.021144229918718338, + 0.021448083221912384, + 0.010541720315814018, + 0.012071170844137669, + 0.011969593353569508, + 0.015338759869337082, + 0.008269356563687325, + 0.01161380484700203, + 0.010534358210861683, + 0.010360941290855408, + 0.006925762165337801, + 0.009545151144266129, + 0.009621204808354378, + 0.010430040769279003, + 0.01710732840001583, + 0.015525507740676403, + 0.01627326011657715, + 0.007847887463867664, + 0.006495613604784012, + 0.012651553377509117, + 0.007021082565188408, + 0.009658445604145527, + 0.02031521685421467, + 0.03965498507022858, + 0.03221341595053673, + 0.05853162705898285, + 0.018964532762765884, + 0.032024044543504715, + 0.030376875773072243, + 0.02945932000875473, + 0.004208545200526714, + 0.007153765764087439, + 0.005666096229106188, + 0.00753592886030674, + 0.00527068180963397, + 0.009108973667025566, + 0.005793847609311342, + 0.006740269251167774, + 0.010726382955908775, + 0.011792533099651337, + 0.008632642216980457, + 0.008962161839008331, + 0.008099897764623165, + 0.012757109478116035, + 0.009902331978082657, + 0.009854613803327084, + 0.007578716147691011, + 0.01065842155367136, + 0.009077238850295544, + 0.009453419595956802, + 0.012691442854702473, + 0.013305897824466228, + 0.008064519613981247, + 0.011468324810266495, + 0.004752360284328461, + 0.0077366274781525135, + 0.006222336087375879, + 0.006881359964609146, + 0.007055756635963917, + 0.0056233275681734085, + 0.004594774916768074, + 0.008912456221878529, + 0.0032228233758360147, + 0.005310173612087965, + 0.00438374187797308, + 0.005402194336056709, + 0.003425989532843232, + 0.0021521581802517176, + 0.0033103672321885824, + 0.0025859647430479527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.13007056713104248 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498329550027847 + ], + "opposite_logprob_tokens": [ + -2.929983377456665 + ], + "logprob_reward": -0.05498329550027847, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498329550027847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14546988904476166, + 0.04101905971765518, + 0.06854000687599182, + 0.19865107536315918, + 0.06485925614833832, + 0.33594924211502075, + 0.06598193198442459, + 0.03741596266627312, + 0.021645061671733856, + 0.020468488335609436 + ], + "credit_map": [ + 0.018791833892464638, + 0.015087341889739037, + 0.017434507608413696, + 0.01442367397248745, + 0.019466467201709747, + 0.012172974646091461, + 0.023045696318149567, + 0.02504740096628666, + 0.005971809849143028, + 0.006083394400775433, + 0.006675523705780506, + 0.005556961987167597, + 0.003650442697107792, + 0.003595462068915367, + 0.004773021209985018, + 0.004712442867457867, + 0.0062494054436683655, + 0.008905438706278801, + 0.00894604530185461, + 0.009143338538706303, + 0.010565745644271374, + 0.007871714420616627, + 0.009526639245450497, + 0.0073316763155162334, + 0.024413008242845535, + 0.02697833813726902, + 0.02991504967212677, + 0.02177245542407036, + 0.019940346479415894, + 0.02456449531018734, + 0.031792495399713516, + 0.019274892285466194, + 0.006437433883547783, + 0.00818269420415163, + 0.00911202747374773, + 0.012769677676260471, + 0.005255761556327343, + 0.007683256641030312, + 0.007546843029558659, + 0.007871564477682114, + 0.02623382769525051, + 0.040113434195518494, + 0.03584237024188042, + 0.05172757804393768, + 0.03543782979249954, + 0.0609079971909523, + 0.04275371879339218, + 0.04293248802423477, + 0.007686879951506853, + 0.011100893840193748, + 0.009283732622861862, + 0.00824813637882471, + 0.005834286101162434, + 0.00955368671566248, + 0.005000791512429714, + 0.00927352998405695, + 0.0029458566568791866, + 0.007908063940703869, + 0.003799837315455079, + 0.003853805363178253, + 0.004158535040915012, + 0.007805352099239826, + 0.0029540511313825846, + 0.003990461118519306, + 0.0017318446189165115, + 0.0025602178648114204, + 0.0024254911113530397, + 0.00258895056322217, + 0.0029675313271582127, + 0.003342483891174197, + 0.0029387602116912603, + 0.003089780919253826, + 0.0017569639021530747, + 0.00364800076931715, + 0.00263356976211071, + 0.0046824561432003975, + 0.0023650575894862413, + 0.001622523064725101, + 0.0022122105583548546, + 0.0015477073611691594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.05498329550027847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620555654168129 + ], + "opposite_logprob_tokens": [ + -2.8120555877685547 + ], + "logprob_reward": -0.0620555654168129, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620555654168129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08174309879541397, + 0.07388664036989212, + 0.07899194210767746, + 0.04092196375131607, + 0.10787943005561829, + 0.16420307755470276, + 0.20782731473445892, + 0.13962814211845398, + 0.07126130163669586, + 0.033657122403383255 + ], + "credit_map": [ + 0.009002824313938618, + 0.01389702595770359, + 0.008527763187885284, + 0.0110460314899683, + 0.01266992837190628, + 0.0061711641028523445, + 0.012039726600050926, + 0.00838862918317318, + 0.012519660405814648, + 0.01286542508751154, + 0.010898416861891747, + 0.008723108097910881, + 0.006742036435753107, + 0.005572743248194456, + 0.007499726954847574, + 0.009065520949661732, + 0.011242130771279335, + 0.009062197990715504, + 0.010445176623761654, + 0.010199674405157566, + 0.008052883669734001, + 0.008957848884165287, + 0.009896752424538136, + 0.011135272681713104, + 0.005713209509849548, + 0.008415157906711102, + 0.005677857901901007, + 0.004000346641987562, + 0.0035610185004770756, + 0.005522695370018482, + 0.003685390343889594, + 0.0043462831526994705, + 0.010021606460213661, + 0.014564020559191704, + 0.01222296617925167, + 0.0182676762342453, + 0.009758266620337963, + 0.012142106890678406, + 0.01610645465552807, + 0.014796322211623192, + 0.014714064076542854, + 0.020333580672740936, + 0.01784021593630314, + 0.02719445526599884, + 0.01801351085305214, + 0.025960523635149002, + 0.021118706092238426, + 0.019028007984161377, + 0.02141740918159485, + 0.0346115306019783, + 0.02426557056605816, + 0.02322814054787159, + 0.026789840310811996, + 0.025391260161995888, + 0.01927371881902218, + 0.032849833369255066, + 0.010114200413227081, + 0.01878662221133709, + 0.018253900110721588, + 0.01626773178577423, + 0.023400600999593735, + 0.023321710526943207, + 0.011879028752446175, + 0.01760433241724968, + 0.004317779093980789, + 0.014854511246085167, + 0.007667898666113615, + 0.007821403443813324, + 0.004676928743720055, + 0.017441516742110252, + 0.006800779141485691, + 0.007680476643145084, + 0.0032704221084713936, + 0.005587321240454912, + 0.004873515572398901, + 0.003699885681271553, + 0.004777370486408472, + 0.003543293569236994, + 0.004275044426321983, + 0.003630265360698104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0620555654168129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369118645787239 + ], + "opposite_logprob_tokens": [ + -3.408691167831421 + ], + "logprob_reward": -0.03369118645787239, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369118645787239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0925171822309494, + 0.04752390831708908, + 0.0920037254691124, + 0.12306736409664154, + 0.3610880672931671, + 0.028631828725337982, + 0.045105427503585815, + 0.09964919835329056, + 0.08513826131820679, + 0.025275053456425667 + ], + "credit_map": [ + 0.008981929160654545, + 0.01442713849246502, + 0.01103928592056036, + 0.012613596394658089, + 0.007718197535723448, + 0.007973678410053253, + 0.020905280485749245, + 0.008858074434101582, + 0.007530183997005224, + 0.0069181183353066444, + 0.006662464700639248, + 0.006670129485428333, + 0.005654344335198402, + 0.0036065757740288973, + 0.00501918513327837, + 0.00546290772035718, + 0.016948388889431953, + 0.014647861942648888, + 0.011793692596256733, + 0.009653558023273945, + 0.008412397466599941, + 0.010964351706206799, + 0.00994693860411644, + 0.00963654089719057, + 0.025156542658805847, + 0.02192399837076664, + 0.020764553919434547, + 0.007704058662056923, + 0.009313371963799, + 0.01698233000934124, + 0.008150354959070683, + 0.013072154484689236, + 0.0409107580780983, + 0.03917747363448143, + 0.044694431126117706, + 0.04979053884744644, + 0.040127649903297424, + 0.04139111191034317, + 0.0629301592707634, + 0.042065925896167755, + 0.002272543963044882, + 0.004939931444823742, + 0.004095533862709999, + 0.0043451678939163685, + 0.0019468640675768256, + 0.0050776503048837185, + 0.0026981146074831486, + 0.0032560229301452637, + 0.004224035888910294, + 0.007438981905579567, + 0.005919455084949732, + 0.005223124288022518, + 0.0037332004867494106, + 0.006920539774000645, + 0.006272778380662203, + 0.005373312160372734, + 0.007017122115939856, + 0.014907656237483025, + 0.01450209878385067, + 0.015511738136410713, + 0.008439202792942524, + 0.018556660041213036, + 0.0075266407802701, + 0.013188081793487072, + 0.0061759804375469685, + 0.009879248216748238, + 0.011752919293940067, + 0.013585840351879597, + 0.009213144890964031, + 0.010282326489686966, + 0.01064314041286707, + 0.013605655170977116, + 0.002872275421395898, + 0.003112443257123232, + 0.004822307266294956, + 0.003973941784352064, + 0.0028288376051932573, + 0.0025146135594695807, + 0.0026275236159563065, + 0.0025231128092855215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.03369118645787239 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798785224556923 + ], + "opposite_logprob_tokens": [ + -3.529798746109009 + ], + "logprob_reward": -0.029798785224556923, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798785224556923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10249847918748856, + 0.03686702996492386, + 0.10669738054275513, + 0.06397486478090286, + 0.11651518195867538, + 0.09453657269477844, + 0.1526593714952469, + 0.15859733521938324, + 0.11683019250631332, + 0.05082361772656441 + ], + "credit_map": [ + 0.017649803310632706, + 0.015372959896922112, + 0.01426861621439457, + 0.013707224279642105, + 0.01175739150494337, + 0.008523676544427872, + 0.008788772858679295, + 0.012430030852556229, + 0.006787317804992199, + 0.005383994895964861, + 0.00505027687177062, + 0.0045959604904055595, + 0.003601966192945838, + 0.0034208297729492188, + 0.003866027109324932, + 0.004160662181675434, + 0.00791615154594183, + 0.013684502802789211, + 0.013823511078953743, + 0.010674766264855862, + 0.016697948798537254, + 0.01311251986771822, + 0.013383244164288044, + 0.017404738813638687, + 0.011827025562524796, + 0.00965668261051178, + 0.011165494099259377, + 0.0047186389565467834, + 0.005508155096322298, + 0.007521434687077999, + 0.0066613806411623955, + 0.006916049402207136, + 0.011748678050935268, + 0.013522624038159847, + 0.014303145930171013, + 0.01960298977792263, + 0.012491157278418541, + 0.01297807041555643, + 0.017621608451008797, + 0.014246908947825432, + 0.006521977484226227, + 0.012923934496939182, + 0.011716756038367748, + 0.01601729355752468, + 0.008651132695376873, + 0.01577727310359478, + 0.010240433737635612, + 0.01268777996301651, + 0.01604088954627514, + 0.02251039631664753, + 0.02083502523601055, + 0.02195562981069088, + 0.011839163489639759, + 0.019642619416117668, + 0.020384974777698517, + 0.01945067197084427, + 0.01134496834129095, + 0.024983573704957962, + 0.013996027410030365, + 0.013327401131391525, + 0.0288770142942667, + 0.02750425413250923, + 0.021012473851442337, + 0.017551634460687637, + 0.008406000211834908, + 0.017104946076869965, + 0.01698986627161503, + 0.01781604066491127, + 0.011501877568662167, + 0.016264773905277252, + 0.013251377269625664, + 0.015495315194129944, + 0.004612758290022612, + 0.0065825218334794044, + 0.005889734718948603, + 0.009467938914895058, + 0.007915666326880455, + 0.006172491237521172, + 0.006352463737130165, + 0.0038300431333482265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.029798785224556923 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle wide shot opens the scene, and the camera begins a slow, deliberate upward tilt, following the momentous ascent of a lone, sleek Martian space elevator pod. As it lifts off, its powerful repulsorgravity engines emit a soft, pulsating blue glow from the vehicle's chassis, silently lifting a vast, swirling cloud of fine ochre dust that billows across the desolate landscape. The pod itself is an aerodynamic teardrop of polished dark gray composite material, its surface unmarked and reflecting the faint sunlight with a subtle sheen. It glides inexorably upward along a vast, shimmering ribbon of semi-translucent material that stretches from the ground to the pale, empty, butterscotch-colored sky, disappearing into the thin upper atmosphere. As the pod climbs higher, its incredibly long, sharp shadow casts a shadow across the cratered terrain below, bisecting the reddish-brown landscape and sweeping across it like the hands of a giant clock. Heavily desaturated color gradients are used throughout key scenes, fading vibrant Martian tones to dusty browns, pale oranges, and cool grays, creating a solemn and awe-inspiring atmosphere that carries profound isolation and historical significance.", + "seed": 29055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_1204_seed_29055_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": -0.10807245448231698, + "vlm_logprob": -0.10807245448231698, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.10807245448231698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11808180809020996 + ], + "opposite_logprob_tokens": [ + -2.24308180809021 + ], + "logprob_reward": -0.11808180809020996, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11808180809020996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14945046603679657, + 0.12478574365377426, + 0.1119493618607521, + 0.041168153285980225, + 0.026398928835988045, + 0.07455792278051376, + 0.103511281311512, + 0.20965512096881866, + 0.1254045069217682, + 0.03311852365732193 + ], + "credit_map": [ + 0.01895746961236, + 0.0151662677526474, + 0.016854438930749893, + 0.01601507142186165, + 0.02306004986166954, + 0.022251784801483154, + 0.024098478257656097, + 0.013046900741755962, + 0.016248593106865883, + 0.013124048709869385, + 0.015734221786260605, + 0.019296886399388313, + 0.009340880438685417, + 0.012590719386935234, + 0.011140864342451096, + 0.02730952762067318, + 0.010121172294020653, + 0.007632258348166943, + 0.017068635672330856, + 0.016043471172451973, + 0.0170444268733263, + 0.013886009342968464, + 0.017070813104510307, + 0.013082575052976608, + 0.004170550499111414, + 0.0027461734134703875, + 0.004692640621215105, + 0.006761377677321434, + 0.003943190909922123, + 0.004309689160436392, + 0.004879414569586515, + 0.009665119461715221, + 0.0029013247694820166, + 0.002675564493983984, + 0.004727297928184271, + 0.0036159399896860123, + 0.0035365596413612366, + 0.0035284599289298058, + 0.0021788040176033974, + 0.0032349780667573214, + 0.010360042564570904, + 0.008995483629405499, + 0.011347214691340923, + 0.011301487684249878, + 0.00601278617978096, + 0.009864846244454384, + 0.006334331352263689, + 0.010341738350689411, + 0.0073105571791529655, + 0.016681905835866928, + 0.014649232849478722, + 0.012672449462115765, + 0.015769198536872864, + 0.015633326023817062, + 0.012556887231767178, + 0.008237726986408234, + 0.012849965132772923, + 0.07213006913661957, + 0.009405015967786312, + 0.011232834309339523, + 0.02320854552090168, + 0.06103109195828438, + 0.010860216803848743, + 0.008937394246459007, + 0.015044746920466423, + 0.024775106459856033, + 0.010977448895573616, + 0.008629354648292065, + 0.02036263421177864, + 0.027344748377799988, + 0.008701595477759838, + 0.009568875655531883, + 0.003201967105269432, + 0.004689713008701801, + 0.004376072436571121, + 0.0035657589323818684, + 0.003802431747317314, + 0.005263537168502808, + 0.0034568195696920156, + 0.0047622243873775005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.11808180809020996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812679275870323 + ], + "opposite_logprob_tokens": [ + -3.2881267070770264 + ], + "logprob_reward": -0.03812679275870323, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812679275870323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058886513113975525, + 0.03394773602485657, + 0.03611036390066147, + 0.03273966535925865, + 0.02823580801486969, + 0.24604006111621857, + 0.06429543346166611, + 0.07187335938215256, + 0.1404748558998108, + 0.2873961925506592 + ], + "credit_map": [ + 0.005259801633656025, + 0.006650872994214296, + 0.009434984996914864, + 0.006307723466306925, + 0.005044623743742704, + 0.009537535719573498, + 0.007759778294712305, + 0.008891190402209759, + 0.004838608670979738, + 0.0028901402838528156, + 0.004776090383529663, + 0.006310374476015568, + 0.0014433355536311865, + 0.004016343969851732, + 0.0026896127965301275, + 0.006983230356127024, + 0.003041038056835532, + 0.0038986175786703825, + 0.0057066152803599834, + 0.007906850427389145, + 0.003139113076031208, + 0.0048621222376823425, + 0.0031792926602065563, + 0.0043767159804701805, + 0.0033528353087604046, + 0.002402642974629998, + 0.003934552893042564, + 0.0052243429236114025, + 0.0022984431125223637, + 0.004509244114160538, + 0.003289937973022461, + 0.007727667223662138, + 0.002638386795297265, + 0.0025547738187015057, + 0.005622100550681353, + 0.004340118262916803, + 0.002980936784297228, + 0.002408175030723214, + 0.003601579926908016, + 0.00408973777666688, + 0.03209514170885086, + 0.02194237895309925, + 0.038814883679151535, + 0.03410155326128006, + 0.031614791601896286, + 0.030922017991542816, + 0.03602276369929314, + 0.02052653208374977, + 0.004703789483755827, + 0.011570666916668415, + 0.008441389538347721, + 0.009423011913895607, + 0.006443461868911982, + 0.009931853972375393, + 0.007262536324560642, + 0.0065187206491827965, + 0.006352569442242384, + 0.0171087384223938, + 0.006127119064331055, + 0.006797458976507187, + 0.008844917640089989, + 0.013105059042572975, + 0.008315442129969597, + 0.005222056992352009, + 0.011139223352074623, + 0.02087375707924366, + 0.013555395416915417, + 0.01387319341301918, + 0.02630004473030567, + 0.02168191596865654, + 0.017805596813559532, + 0.015245726332068443, + 0.023544171825051308, + 0.06078578159213066, + 0.031254347413778305, + 0.023323001340031624, + 0.03084941953420639, + 0.06525718420743942, + 0.030096791684627533, + 0.022285480052232742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.03812679275870323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210386514663696 + ], + "opposite_logprob_tokens": [ + -1.5021038055419922 + ], + "logprob_reward": -0.25210386514663696, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210386514663696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08218444138765335, + 0.043103981763124466, + 0.022695984691381454, + 0.030326949432492256, + 0.03861518204212189, + 0.33070871233940125, + 0.07905064523220062, + 0.23511728644371033, + 0.04694381356239319, + 0.09125302731990814 + ], + "credit_map": [ + 0.008794335648417473, + 0.009267987683415413, + 0.011234717443585396, + 0.00864387210458517, + 0.009198091924190521, + 0.014770494773983955, + 0.009917281568050385, + 0.010357655584812164, + 0.005470366682857275, + 0.0062963105738162994, + 0.0051430705934762955, + 0.008509271778166294, + 0.002859108615666628, + 0.00604398874565959, + 0.003655946347862482, + 0.005125915631651878, + 0.0028236520010977983, + 0.0021597493905574083, + 0.002972652204334736, + 0.00359780783765018, + 0.0023464581463485956, + 0.0030537263955920935, + 0.00239715282805264, + 0.003344782628118992, + 0.002917726757004857, + 0.0024733594618737698, + 0.004032347351312637, + 0.003912276588380337, + 0.003879451658576727, + 0.0038584035355597734, + 0.005215540993958712, + 0.004037841223180294, + 0.0036533039528876543, + 0.0033177374862134457, + 0.005000242032110691, + 0.006123471073806286, + 0.0033797030337154865, + 0.004680688958615065, + 0.00346160214394331, + 0.008998431265354156, + 0.033572420477867126, + 0.03566790372133255, + 0.04005838930606842, + 0.04380915313959122, + 0.042414650321006775, + 0.05750221386551857, + 0.0404764860868454, + 0.037207476794719696, + 0.006860520225018263, + 0.009650168009102345, + 0.012494873255491257, + 0.011014344170689583, + 0.011802742257714272, + 0.012587581761181355, + 0.007435524836182594, + 0.007204888854175806, + 0.017132284119725227, + 0.07339367270469666, + 0.020498689264059067, + 0.01993931271135807, + 0.016852622851729393, + 0.058096036314964294, + 0.010456724092364311, + 0.01874794438481331, + 0.005653655156493187, + 0.00699066324159503, + 0.005907727405428886, + 0.0051146638579666615, + 0.005697032902389765, + 0.006835775449872017, + 0.005551553796976805, + 0.005192737095057964, + 0.007980919443070889, + 0.014996010810136795, + 0.010411759838461876, + 0.00785453338176012, + 0.014835821464657784, + 0.017072608694434166, + 0.00683590117841959, + 0.011265460401773453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.25210386514663696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700092539191246 + ], + "opposite_logprob_tokens": [ + -2.695009231567383 + ], + "logprob_reward": -0.0700092539191246, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700092539191246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13294750452041626, + 0.1645030975341797, + 0.04494465887546539, + 0.03393390402197838, + 0.016084926202893257, + 0.06408533453941345, + 0.08168769627809525, + 0.18087682127952576, + 0.25427931547164917, + 0.026656774803996086 + ], + "credit_map": [ + 0.016850553452968597, + 0.011805753223598003, + 0.016965100541710854, + 0.014626740477979183, + 0.018261432647705078, + 0.02000400796532631, + 0.02194676548242569, + 0.012487150728702545, + 0.025057600811123848, + 0.018873771652579308, + 0.022464891895651817, + 0.02745707705616951, + 0.009207508526742458, + 0.016836782917380333, + 0.01160274539142847, + 0.03300271928310394, + 0.003796262899413705, + 0.00381539692170918, + 0.009408364072442055, + 0.007397830951958895, + 0.0051809088326990604, + 0.004685938358306885, + 0.004587723407894373, + 0.006072236690670252, + 0.003267768071964383, + 0.0033747516572475433, + 0.00543781416490674, + 0.004165124148130417, + 0.005658873822540045, + 0.003162484848871827, + 0.005394095554947853, + 0.003472990123555064, + 0.0018838882679119706, + 0.0019399333978071809, + 0.0021337245125323534, + 0.002248542383313179, + 0.0016930095152929425, + 0.0016740461578592658, + 0.0022391323000192642, + 0.0022726499009877443, + 0.004654189106076956, + 0.0077426196075975895, + 0.010438630357384682, + 0.009142961353063583, + 0.009023718535900116, + 0.006238929461687803, + 0.010195584036409855, + 0.006648701149970293, + 0.004339155741035938, + 0.009377283975481987, + 0.013347836211323738, + 0.011942124925553799, + 0.010433748364448547, + 0.007772322744131088, + 0.013890051282942295, + 0.010585164651274681, + 0.01176425814628601, + 0.06426525115966797, + 0.013549906201660633, + 0.011873285286128521, + 0.011020085774362087, + 0.049716074019670486, + 0.008063052780926228, + 0.010624893009662628, + 0.02650039829313755, + 0.03550826385617256, + 0.028061656281352043, + 0.029353367164731026, + 0.03817260265350342, + 0.03782946616411209, + 0.028818678110837936, + 0.03003489226102829, + 0.0021928525529801846, + 0.007263106293976307, + 0.002383164595812559, + 0.002266440773382783, + 0.00209583411924541, + 0.007227962836623192, + 0.001156059093773365, + 0.0020713561680167913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.0700092539191246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062040552496910095 + ], + "opposite_logprob_tokens": [ + -2.8120405673980713 + ], + "logprob_reward": -0.062040552496910095, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062040552496910095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07526043802499771, + 0.049466148018836975, + 0.04168839752674103, + 0.08138115704059601, + 0.05879933387041092, + 0.26184555888175964, + 0.1258455216884613, + 0.1361882984638214, + 0.07496238499879837, + 0.09456281363964081 + ], + "credit_map": [ + 0.006169045809656382, + 0.008107857778668404, + 0.009022686630487442, + 0.008507853373885155, + 0.007877986878156662, + 0.012188761495053768, + 0.011621523648500443, + 0.011764714494347572, + 0.005233069881796837, + 0.007560212165117264, + 0.0062827761285007, + 0.007679678034037352, + 0.004874803125858307, + 0.007543791551142931, + 0.005328210070729256, + 0.00496360007673502, + 0.003757682628929615, + 0.004720652941614389, + 0.0056977346539497375, + 0.006841759663075209, + 0.004864661023020744, + 0.005582146346569061, + 0.0048357294872403145, + 0.005388029385358095, + 0.006701657082885504, + 0.006008438300341368, + 0.01115467119961977, + 0.010752376168966293, + 0.010952056385576725, + 0.008325898088514805, + 0.017403705045580864, + 0.010082351043820381, + 0.004918292164802551, + 0.005120504647493362, + 0.01183933112770319, + 0.009480448439717293, + 0.005190850701183081, + 0.006066227797418833, + 0.007774872239679098, + 0.008408796042203903, + 0.02004936896264553, + 0.023230131715536118, + 0.040524858981370926, + 0.036804504692554474, + 0.03654524311423302, + 0.0305678341537714, + 0.0417281836271286, + 0.03239541873335838, + 0.010240728966891766, + 0.019218433648347855, + 0.02270534448325634, + 0.018161015585064888, + 0.016564829275012016, + 0.013345992192626, + 0.010563012212514877, + 0.015046165324747562, + 0.014691579155623913, + 0.015203773975372314, + 0.024702688679099083, + 0.01953580044209957, + 0.01450230274349451, + 0.018172334879636765, + 0.014619268476963043, + 0.014760549180209637, + 0.005562159698456526, + 0.010204225778579712, + 0.00703765032812953, + 0.008845210075378418, + 0.0130815664306283, + 0.010903647169470787, + 0.00847857166081667, + 0.010849345475435257, + 0.008587134070694447, + 0.025447668507695198, + 0.007115521933883429, + 0.007092148531228304, + 0.006749295629560947, + 0.02704627625644207, + 0.00402187742292881, + 0.008502879180014133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.062040552496910095 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle wide shot opens the scene, and the camera begins a slow, deliberate upward tilt, following the momentous ascent of a lone, sleek Martian space elevator pod. As it lifts off, its powerful repulsorgravity engines emit a soft, pulsating blue glow from the vehicle's chassis, silently lifting a vast, swirling cloud of fine ochre dust that billows across the desolate landscape. The pod itself is an aerodynamic teardrop of polished dark gray composite material, its surface unmarked and reflecting the faint sunlight with a subtle sheen. It glides inexorably upward along a vast, shimmering ribbon of semi-translucent material that stretches from the ground to the pale, empty, butterscotch-colored sky, disappearing into the thin upper atmosphere. As the pod climbs higher, its incredibly long, sharp shadow casts a shadow across the cratered terrain below, bisecting the reddish-brown landscape and sweeping across it like the hands of a giant clock. Heavily desaturated color gradients are used throughout key scenes, fading vibrant Martian tones to dusty browns, pale oranges, and cool grays, creating a solemn and awe-inspiring atmosphere that carries profound isolation and historical significance.", + "seed": 29056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_1204_seed_29056_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": -0.06838786378502845, + "vlm_logprob": -0.06838786378502845, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.06838786378502845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03740452602505684 + ], + "opposite_logprob_tokens": [ + -3.6624045372009277 + ], + "logprob_reward": -0.03740452602505684, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03740452602505684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08662679046392441, + 0.02600243128836155, + 0.04121461883187294, + 0.017322901636362076, + 0.015572465024888515, + 0.020377758890390396, + 0.021222226321697235, + 0.7236810326576233, + 0.028246473520994186, + 0.019733330234885216 + ], + "credit_map": [ + 0.005257124546915293, + 0.01049251202493906, + 0.010725492611527443, + 0.017286401242017746, + 0.01029192004352808, + 0.0063081709668040276, + 0.011520370841026306, + 0.014744803309440613, + 0.0023853296879678965, + 0.00449790945276618, + 0.00271398201584816, + 0.0025138589553534985, + 0.0034797736443579197, + 0.0034612121526151896, + 0.0042491392232477665, + 0.002701226156204939, + 0.0026429020799696445, + 0.006531048566102982, + 0.00317402184009552, + 0.004587571136653423, + 0.006882404442876577, + 0.006988592445850372, + 0.006569717079401016, + 0.0038383598439395428, + 0.001382439280860126, + 0.0024166994262486696, + 0.0012618839973583817, + 0.0031061640474945307, + 0.0022052484564483166, + 0.0025358835700899363, + 0.002367790089920163, + 0.0020467930007725954, + 0.001890460029244423, + 0.0024679203052073717, + 0.0012495005503296852, + 0.0018111513927578926, + 0.0017391244182363153, + 0.0029413707088679075, + 0.0015255825128406286, + 0.00194735627155751, + 0.004551072139292955, + 0.0037519370671361685, + 0.001989947631955147, + 0.0024077785201370716, + 0.0016433591954410076, + 0.0023388105910271406, + 0.0017446748679503798, + 0.001950178062543273, + 0.001735497498884797, + 0.002795337000861764, + 0.0043281144462525845, + 0.0036918993573635817, + 0.001749734510667622, + 0.0024744130205363035, + 0.002013124292716384, + 0.002434107707813382, + 0.03562239557504654, + 0.0786433219909668, + 0.23217278718948364, + 0.13839411735534668, + 0.01597634330391884, + 0.14748568832874298, + 0.014922614209353924, + 0.060463711619377136, + 0.0025132226292043924, + 0.004790144506841898, + 0.00347484671510756, + 0.003650365397334099, + 0.0022659790702164173, + 0.003364631673321128, + 0.002321788342669606, + 0.005865495651960373, + 0.0011828087735921144, + 0.0016675125807523727, + 0.003018860472366214, + 0.003006912535056472, + 0.003079734742641449, + 0.0022969383280724287, + 0.002944278996437788, + 0.0025362831074744463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.03740452602505684 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054980698972940445 + ], + "opposite_logprob_tokens": [ + -2.929980754852295 + ], + "logprob_reward": -0.054980698972940445, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054980698972940445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10836918652057648, + 0.016096333041787148, + 0.023320740088820457, + 0.011419716291129589, + 0.009307557716965675, + 0.018508801236748695, + 0.00876529049128294, + 0.7814821600914001, + 0.011253357864916325, + 0.011476901359856129 + ], + "credit_map": [ + 0.009869372472167015, + 0.010790014639496803, + 0.018834764137864113, + 0.019952859729528427, + 0.01225773524492979, + 0.008800337091088295, + 0.017618294805288315, + 0.010245799086987972, + 0.0015457305125892162, + 0.002366460394114256, + 0.0012006021570414305, + 0.001456861267797649, + 0.0022330982610583305, + 0.00228860997594893, + 0.002894190838560462, + 0.0021107776556164026, + 0.0014244532212615013, + 0.003222663188353181, + 0.001610104343853891, + 0.0024391571059823036, + 0.0036796550266444683, + 0.003951431252062321, + 0.004599686712026596, + 0.0023935846984386444, + 0.0011088945902884007, + 0.0028022327460348606, + 0.0009149075485765934, + 0.0014158831909298897, + 0.001090227160602808, + 0.001609268132597208, + 0.0013829013332724571, + 0.0010954000754281878, + 0.0008971889619715512, + 0.001256613410077989, + 0.0007143416441977024, + 0.0011857133358716965, + 0.001528318622149527, + 0.0011069095926359296, + 0.0017468577716499567, + 0.0008716135052964091, + 0.0029968563467264175, + 0.003287568688392639, + 0.002087041735649109, + 0.002270875731483102, + 0.0018436760874465108, + 0.0025618476793169975, + 0.0013782430905848742, + 0.0020826912950724363, + 0.0007036504684947431, + 0.001273650792427361, + 0.001495607546530664, + 0.001560723758302629, + 0.0008476582588627934, + 0.0008790866122581065, + 0.001043697469867766, + 0.0009612152935005724, + 0.039362676441669464, + 0.08494017273187637, + 0.2507821321487427, + 0.14998763799667358, + 0.015906762331724167, + 0.16047216951847076, + 0.014418274164199829, + 0.06561227142810822, + 0.0007559219375252724, + 0.0011024075793102384, + 0.0016271689673885703, + 0.001578127732500434, + 0.0015541415195912123, + 0.0009963997872546315, + 0.0016937763430178165, + 0.001945412135683, + 0.0008349061827175319, + 0.0010063725057989359, + 0.0012312986655160785, + 0.0015177377499639988, + 0.00189150997903198, + 0.0012313027400523424, + 0.0026954992208629847, + 0.0010682727443054318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.054980698972940445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015710175037384 + ], + "opposite_logprob_tokens": [ + -1.701570987701416 + ], + "logprob_reward": -0.2015710175037384, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015710175037384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1187901645898819, + 0.021061604842543602, + 0.025707410648465157, + 0.024362364783883095, + 0.03799399733543396, + 0.030642595142126083, + 0.015098422765731812, + 0.6832315325737, + 0.019893594086170197, + 0.02321825735270977 + ], + "credit_map": [ + 0.005962443072348833, + 0.008955233730375767, + 0.012237266637384892, + 0.030855976045131683, + 0.006227244157344103, + 0.013048677705228329, + 0.010687675327062607, + 0.030815649777650833, + 0.0021570981480181217, + 0.003521902719512582, + 0.0019150302978232503, + 0.0018924257019534707, + 0.0026877252385020256, + 0.0027890733908861876, + 0.003912623040378094, + 0.002185726072639227, + 0.0021608483511954546, + 0.004392251372337341, + 0.0023854169994592667, + 0.004557470791041851, + 0.0023614850360900164, + 0.0047151329927146435, + 0.0026087414007633924, + 0.0025260625407099724, + 0.002712382236495614, + 0.007770972326397896, + 0.0017254238482564688, + 0.00276587693952024, + 0.0014011407038196921, + 0.004081449471414089, + 0.001406357972882688, + 0.0024987603537738323, + 0.005240915343165398, + 0.005137286614626646, + 0.004295036196708679, + 0.0033934933599084616, + 0.0017813703743740916, + 0.008997596800327301, + 0.0017746826633810997, + 0.007373617962002754, + 0.00512969633564353, + 0.004754253663122654, + 0.0033685436937958, + 0.003953525330871344, + 0.0026857333723455667, + 0.005370688624680042, + 0.0017116664675995708, + 0.0036684884689748287, + 0.0012096583377569914, + 0.0016252686036750674, + 0.002179643837735057, + 0.001955298474058509, + 0.002362848725169897, + 0.0015252061421051621, + 0.0021547020878642797, + 0.002085797255858779, + 0.0337306372821331, + 0.07442950457334518, + 0.21979089081287384, + 0.1306980550289154, + 0.014196408912539482, + 0.13980010151863098, + 0.013144949451088905, + 0.05744100734591484, + 0.001736617530696094, + 0.001932449173182249, + 0.0031810696236789227, + 0.002710975706577301, + 0.00203928304836154, + 0.0023093370255082846, + 0.001799559686332941, + 0.004184300545603037, + 0.0019640543032437563, + 0.0028246697038412094, + 0.002841173205524683, + 0.0025271137710660696, + 0.002217329340055585, + 0.003983463626354933, + 0.004486532881855965, + 0.0023739186581224203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.2015710175037384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796123504638672 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.029796123504638672, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796123504638672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023203305900096893, + 0.022525301203131676, + 0.014990478754043579, + 0.012426773086190224, + 0.017366237938404083, + 0.014516397379338741, + 0.02042585052549839, + 0.8295078277587891, + 0.03221052139997482, + 0.012827272526919842 + ], + "credit_map": [ + 0.0015159512404352427, + 0.002509190933778882, + 0.0027850153855979443, + 0.005712222773581743, + 0.0015755584463477135, + 0.0023697903379797935, + 0.0021498752757906914, + 0.004585702437907457, + 0.0021628718823194504, + 0.0029938602820038795, + 0.002465131925418973, + 0.0020670052617788315, + 0.003404969349503517, + 0.0030958729330450296, + 0.004072519950568676, + 0.0022630703169852495, + 0.0010226970771327615, + 0.0024485986214131117, + 0.0009489385993219912, + 0.0015642859507352114, + 0.002178834518417716, + 0.0025014893617480993, + 0.002353225601837039, + 0.00197241036221385, + 0.0012949249939993024, + 0.002753898734226823, + 0.0011032259790226817, + 0.0016889419639483094, + 0.0010574277257546782, + 0.001626852317713201, + 0.0017584494780749083, + 0.001143052475526929, + 0.002072773175314069, + 0.0037428010255098343, + 0.0017581043066456914, + 0.0016495737945660949, + 0.0016398006118834019, + 0.0026036968920379877, + 0.001976755680516362, + 0.0019227330340072513, + 0.002094671130180359, + 0.002305365400388837, + 0.0014020674861967564, + 0.0019032772397622466, + 0.0016303259180858731, + 0.00177475914824754, + 0.0014099464751780033, + 0.0019959856290370226, + 0.0017674299888312817, + 0.0032460198272019625, + 0.003334187902510166, + 0.003145491471514106, + 0.0028095918241888285, + 0.0023979898542165756, + 0.0017481077229604125, + 0.0019770346116274595, + 0.041919026523828506, + 0.08974657952785492, + 0.26626500487327576, + 0.15956942737102509, + 0.016849888488650322, + 0.17007645964622498, + 0.01556875929236412, + 0.0695127621293068, + 0.0023008445277810097, + 0.0039254589937627316, + 0.004715630318969488, + 0.005149921867996454, + 0.0023357891477644444, + 0.0029895103070884943, + 0.0037675178609788418, + 0.0070258500054478645, + 0.0009860513964667916, + 0.0010618316009640694, + 0.0012284782715141773, + 0.0013660900294780731, + 0.0012225379468873143, + 0.0015244422247633338, + 0.003760027466341853, + 0.0016778138233348727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.029796123504638672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818695291876793 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818695291876793, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818695291876793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052709899842739105, + 0.00960824079811573, + 0.009234716184437275, + 0.004894549492746592, + 0.007219433784484863, + 0.011408969759941101, + 0.023269427940249443, + 0.8348523378372192, + 0.015673013404011726, + 0.031129322946071625 + ], + "credit_map": [ + 0.005205592606216669, + 0.005420341622084379, + 0.007608900777995586, + 0.010827014222741127, + 0.004638150334358215, + 0.00451273238286376, + 0.004218601621687412, + 0.01027856208384037, + 0.0007371308165602386, + 0.0016671536723151803, + 0.0009116443106904626, + 0.0010320849251002073, + 0.001131138182245195, + 0.0014101312262937427, + 0.00187157210893929, + 0.0008473857305943966, + 0.0008787799743004143, + 0.0014390656724572182, + 0.0010664649307727814, + 0.001132854144088924, + 0.0006407466717064381, + 0.0016607638681307435, + 0.001073450199328363, + 0.0013425910146906972, + 0.000547965697478503, + 0.001054172869771719, + 0.00048771873116493225, + 0.0006183881778270006, + 0.00040530195110477507, + 0.0006950651877559721, + 0.0005366555415093899, + 0.0005492811906151474, + 0.0008226818172261119, + 0.0011833157623186707, + 0.0008300107438117266, + 0.0010704047745093703, + 0.0006372826755978167, + 0.0008094232180155814, + 0.0007264298619702458, + 0.0011398849310353398, + 0.0016320307040587068, + 0.0013255274388939142, + 0.001520973164588213, + 0.0019348927307873964, + 0.0009412535582669079, + 0.0018911582883447409, + 0.0009640222415328026, + 0.001199111808091402, + 0.0015807409072294831, + 0.002901668893173337, + 0.004052548203617334, + 0.0037615199107676744, + 0.003484014654532075, + 0.0027914519887417555, + 0.002542097121477127, + 0.0021553863771259785, + 0.041709642857313156, + 0.09043128043413162, + 0.2679973840713501, + 0.16059711575508118, + 0.017547810450196266, + 0.1710333377122879, + 0.015180186368525028, + 0.07035563141107559, + 0.0012468235800042748, + 0.002127567306160927, + 0.002740270923823118, + 0.002287904964759946, + 0.0011239979648962617, + 0.002373649040237069, + 0.002056579804047942, + 0.0017162186559289694, + 0.002341552870348096, + 0.0025920141488313675, + 0.004380541853606701, + 0.004562440328299999, + 0.0022395923733711243, + 0.0036127271596342325, + 0.007664483971893787, + 0.0037359707057476044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.01818695291876793 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle wide shot opens the scene, and the camera begins a slow, deliberate upward tilt, following the momentous ascent of a lone, sleek Martian space elevator pod. As it lifts off, its powerful repulsorgravity engines emit a soft, pulsating blue glow from the vehicle's chassis, silently lifting a vast, swirling cloud of fine ochre dust that billows across the desolate landscape. The pod itself is an aerodynamic teardrop of polished dark gray composite material, its surface unmarked and reflecting the faint sunlight with a subtle sheen. It glides inexorably upward along a vast, shimmering ribbon of semi-translucent material that stretches from the ground to the pale, empty, butterscotch-colored sky, disappearing into the thin upper atmosphere. As the pod climbs higher, its incredibly long, sharp shadow casts a shadow across the cratered terrain below, bisecting the reddish-brown landscape and sweeping across it like the hands of a giant clock. Heavily desaturated color gradients are used throughout key scenes, fading vibrant Martian tones to dusty browns, pale oranges, and cool grays, creating a solemn and awe-inspiring atmosphere that carries profound isolation and historical significance.", + "seed": 29057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_13/index_1204_seed_29057_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": -0.09132572114467621, + "vlm_logprob": -0.09132572114467621, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.09132572114467621, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07591060549020767 + ], + "opposite_logprob_tokens": [ + -2.7009105682373047 + ], + "logprob_reward": -0.07591060549020767, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07591060549020767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0895332396030426, + 0.03734507039189339, + 0.049765221774578094, + 0.13464900851249695, + 0.054945625364780426, + 0.02200121060013771, + 0.4461427628993988, + 0.10995947569608688, + 0.03179879114031792, + 0.02385961450636387 + ], + "credit_map": [ + 0.008444171398878098, + 0.009445572271943092, + 0.015079394914209843, + 0.010915852151811123, + 0.011525494046509266, + 0.01357046328485012, + 0.010955807752907276, + 0.009596478193998337, + 0.004850670695304871, + 0.0034360350109636784, + 0.004171150736510754, + 0.0051965550519526005, + 0.0065591055899858475, + 0.005209174007177353, + 0.005011586006730795, + 0.002910797717049718, + 0.0048629045486450195, + 0.005910394713282585, + 0.005538043100386858, + 0.009150288067758083, + 0.006933707743883133, + 0.006868923548609018, + 0.0059935953468084335, + 0.004507363308221102, + 0.013125244528055191, + 0.01750539056956768, + 0.01772020384669304, + 0.02097358927130699, + 0.017990732565522194, + 0.017257660627365112, + 0.013547186739742756, + 0.016529008746147156, + 0.00415416993200779, + 0.010566589422523975, + 0.0051532224752008915, + 0.008236751891672611, + 0.006120976060628891, + 0.008654418401420116, + 0.003874802263453603, + 0.008184695616364479, + 0.002006041118875146, + 0.0023088084999471903, + 0.002923200372606516, + 0.0039215171709656715, + 0.0026785589288920164, + 0.0026154140941798687, + 0.0029196294490247965, + 0.002628039103001356, + 0.03522645682096481, + 0.07343477755784988, + 0.047423820942640305, + 0.04717157036066055, + 0.07879652827978134, + 0.05850328132510185, + 0.04251030460000038, + 0.06307601928710938, + 0.01132905576378107, + 0.019399791955947876, + 0.006262829527258873, + 0.006911534816026688, + 0.02179567515850067, + 0.022852007299661636, + 0.010044952854514122, + 0.011363619938492775, + 0.003645433345809579, + 0.005856912117451429, + 0.0027184244245290756, + 0.0030287993140518665, + 0.0037147740367799997, + 0.004229161888360977, + 0.0032053415197879076, + 0.0053999414667487144, + 0.002497589448466897, + 0.003725868184119463, + 0.0024118241854012012, + 0.0036474387161433697, + 0.0026996182277798653, + 0.00326986494474113, + 0.003875579684972763, + 0.0017318306490778923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.07591060549020767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309525713324547 + ], + "opposite_logprob_tokens": [ + -3.168095350265503 + ], + "logprob_reward": -0.04309525713324547, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309525713324547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30625343322753906, + 0.18265506625175476, + 0.0715307667851448, + 0.05144388601183891, + 0.03687280789017677, + 0.05581928417086601, + 0.1892736852169037, + 0.038482118397951126, + 0.0429917648434639, + 0.024677230045199394 + ], + "credit_map": [ + 0.025120239704847336, + 0.03008241020143032, + 0.021935952827334404, + 0.032172415405511856, + 0.09470445662736893, + 0.047932107001543045, + 0.03575873002409935, + 0.018547112122178078, + 0.01870114356279373, + 0.021459849551320076, + 0.02049332484602928, + 0.02244807407259941, + 0.033285897225141525, + 0.024597452953457832, + 0.021514175459742546, + 0.02015514485538006, + 0.007906039245426655, + 0.00787806324660778, + 0.0066701490432024, + 0.00986342504620552, + 0.0111091248691082, + 0.009598024189472198, + 0.009751896373927593, + 0.008754041977226734, + 0.004903229884803295, + 0.005450898315757513, + 0.006609585136175156, + 0.007979746907949448, + 0.00693866703659296, + 0.005840274039655924, + 0.007281721569597721, + 0.006439762655645609, + 0.0029792783316224813, + 0.005326363258063793, + 0.005347897298634052, + 0.004939613863825798, + 0.0036865100264549255, + 0.007176810875535011, + 0.0027553332038223743, + 0.004661000799387693, + 0.00484482804313302, + 0.006708355154842138, + 0.005863326136022806, + 0.008355338126420975, + 0.009685379453003407, + 0.0061480761505663395, + 0.006821158342063427, + 0.007392823696136475, + 0.019508270546793938, + 0.03165709599852562, + 0.02090483158826828, + 0.02058487758040428, + 0.02998693659901619, + 0.02720593474805355, + 0.014072097837924957, + 0.025353634729981422, + 0.004477477632462978, + 0.006396119017153978, + 0.0037193112075328827, + 0.0035819734912365675, + 0.00618424266576767, + 0.0050031584687530994, + 0.004464981146156788, + 0.004654855467379093, + 0.0038299669977277517, + 0.005371773149818182, + 0.0038793699350208044, + 0.00728081027045846, + 0.005748635157942772, + 0.004094536881893873, + 0.0050820112228393555, + 0.0077046602964401245, + 0.002829184290021658, + 0.004021462053060532, + 0.002119028475135565, + 0.003663642331957817, + 0.0026833992451429367, + 0.0030765626579523087, + 0.0023404075764119625, + 0.0039435443468391895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04309525713324547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813430428504944 + ], + "opposite_logprob_tokens": [ + -1.4063429832458496 + ], + "logprob_reward": -0.2813430428504944, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813430428504944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1249965950846672, + 0.0827740728855133, + 0.07875098288059235, + 0.08124663680791855, + 0.07157488167285919, + 0.06533082574605942, + 0.31809502840042114, + 0.07974826544523239, + 0.07520356774330139, + 0.022279130294919014 + ], + "credit_map": [ + 0.011273396201431751, + 0.014052865095436573, + 0.017280051484704018, + 0.01906832493841648, + 0.019814563915133476, + 0.016371723264455795, + 0.015859974548220634, + 0.011275693774223328, + 0.008461098186671734, + 0.008832462131977081, + 0.011206821538507938, + 0.016832483932375908, + 0.010884055867791176, + 0.011292909272015095, + 0.007932332344353199, + 0.007331905420869589, + 0.011810097843408585, + 0.010525225661695004, + 0.010263689793646336, + 0.015392059460282326, + 0.0057403575628995895, + 0.010721531696617603, + 0.006162741221487522, + 0.008135278709232807, + 0.008824868127703667, + 0.009996065869927406, + 0.01080385036766529, + 0.0140566760674119, + 0.010911323130130768, + 0.008135301992297173, + 0.008787209168076515, + 0.009731339290738106, + 0.0053986418060958385, + 0.009011982940137386, + 0.012410818599164486, + 0.01008797436952591, + 0.008011667989194393, + 0.009471898898482323, + 0.007381046656519175, + 0.009800850413739681, + 0.005012743640691042, + 0.007986507378518581, + 0.007583229802548885, + 0.009813101962208748, + 0.009565266780555248, + 0.007294126786291599, + 0.009347345679998398, + 0.008728502318263054, + 0.026095286011695862, + 0.05127766355872154, + 0.03557952120900154, + 0.03604242205619812, + 0.05180470272898674, + 0.04347119852900505, + 0.025980228558182716, + 0.047844015061855316, + 0.00867683719843626, + 0.011647904291749, + 0.006505013909190893, + 0.007164031732827425, + 0.01762404851615429, + 0.008598621003329754, + 0.006516596302390099, + 0.013015210628509521, + 0.008084111846983433, + 0.019519539549946785, + 0.004202594514936209, + 0.009371611289680004, + 0.005271512549370527, + 0.013047563843429089, + 0.004568943753838539, + 0.011137695983052254, + 0.0015894531970843673, + 0.002843274734914303, + 0.003565050894394517, + 0.005067688878625631, + 0.0024495795369148254, + 0.0019235105719417334, + 0.00218894868157804, + 0.002651624148711562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.2813430428504944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809661045670509 + ], + "opposite_logprob_tokens": [ + -3.2880966663360596 + ], + "logprob_reward": -0.03809661045670509, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809661045670509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12546727061271667, + 0.07323972135782242, + 0.12418659776449203, + 0.13358111679553986, + 0.03917791694402695, + 0.16632039844989777, + 0.15241527557373047, + 0.10128925740718842, + 0.05006653070449829, + 0.03425593301653862 + ], + "credit_map": [ + 0.01130341924726963, + 0.015032747760415077, + 0.018036065623164177, + 0.014411300420761108, + 0.01898930035531521, + 0.016546016559004784, + 0.01676500216126442, + 0.014383415691554546, + 0.008601894602179527, + 0.0077208527363836765, + 0.0076293619349598885, + 0.009767604060471058, + 0.01302478089928627, + 0.01042408961802721, + 0.009477001614868641, + 0.006594136822968721, + 0.014717516489326954, + 0.015360384248197079, + 0.0141462916508317, + 0.023052005097270012, + 0.014251358807086945, + 0.01642032526433468, + 0.014734694734215736, + 0.011504016816616058, + 0.012814603745937347, + 0.012423010542988777, + 0.020754965022206306, + 0.021899845451116562, + 0.015132322907447815, + 0.01533007062971592, + 0.01505938172340393, + 0.020166922360658646, + 0.0029041667003184557, + 0.0052196611650288105, + 0.005472172982990742, + 0.00657085794955492, + 0.004297382198274136, + 0.006147869396954775, + 0.0034157512709498405, + 0.005150051787495613, + 0.011500944383442402, + 0.020716281607747078, + 0.019965115934610367, + 0.026143677532672882, + 0.02221011556684971, + 0.020251082256436348, + 0.023589687421917915, + 0.02194349095225334, + 0.012900802306830883, + 0.024219227954745293, + 0.016569191589951515, + 0.015731830149888992, + 0.02886599861085415, + 0.019750965759158134, + 0.011342053301632404, + 0.023035205900669098, + 0.007507896516472101, + 0.022987807169556618, + 0.0060506947338581085, + 0.007556406781077385, + 0.015705779194831848, + 0.022943904623389244, + 0.008580805733799934, + 0.009955955669283867, + 0.006727215833961964, + 0.009019717574119568, + 0.004256691318005323, + 0.0068258261308074, + 0.004358801990747452, + 0.0061469110660254955, + 0.0035629926715046167, + 0.00916837528347969, + 0.003579485695809126, + 0.005140690598636866, + 0.004424293525516987, + 0.006606611423194408, + 0.0032211397774517536, + 0.003441023640334606, + 0.0036170906387269497, + 0.004225597251206636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.03809661045670509 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the pod shaped like an aerodynamic teardrop made of polished dark gray composite material?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the pod's repulsorgravity engines emit a soft, pulsating blue glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pod lift off while silently billowing a vast, swirling cloud of fine ochre dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera starting with a low-angle wide shot and performing a slow, deliberate upward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the pod gliding upward along a vast, shimmering ribbon of semi-translucent material stretching to the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183089792728424 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183089792728424, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183089792728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06690504401922226, + 0.04841543734073639, + 0.10919231921434402, + 0.07518957555294037, + 0.06444226205348969, + 0.19955329596996307, + 0.2066965252161026, + 0.04661791771650314, + 0.10955436527729034, + 0.0734333023428917 + ], + "credit_map": [ + 0.00705867912620306, + 0.007216468453407288, + 0.011733313091099262, + 0.009636851958930492, + 0.009545136243104935, + 0.009536479599773884, + 0.007466391660273075, + 0.00471171410754323, + 0.005718458443880081, + 0.005789503455162048, + 0.007840326987206936, + 0.006805403623729944, + 0.008248885162174702, + 0.006223464850336313, + 0.004873455967754126, + 0.002915932098403573, + 0.016385255381464958, + 0.01381929125636816, + 0.013505566865205765, + 0.02087252028286457, + 0.00804656557738781, + 0.014397785067558289, + 0.010097122751176357, + 0.012068199925124645, + 0.007747068535536528, + 0.008199155330657959, + 0.015880294144153595, + 0.011410118080675602, + 0.007734784856438637, + 0.008035057224333286, + 0.005930606741458178, + 0.010252476669847965, + 0.005267976783216, + 0.01136177871376276, + 0.008245163597166538, + 0.011126180179417133, + 0.006464533507823944, + 0.008010723628103733, + 0.0038197077810764313, + 0.010146195068955421, + 0.013140085153281689, + 0.02714608423411846, + 0.02555878832936287, + 0.033775731921195984, + 0.03050202876329422, + 0.02518896944820881, + 0.018143609166145325, + 0.026097966358065605, + 0.02065681293606758, + 0.035963401198387146, + 0.0216275155544281, + 0.020775701850652695, + 0.03169459477066994, + 0.03210480511188507, + 0.014921411871910095, + 0.02895227074623108, + 0.005592857953161001, + 0.009207637049257755, + 0.0031214032787829638, + 0.0031640862580388784, + 0.0064177862368524075, + 0.010919170454144478, + 0.0028227302245795727, + 0.005372237879782915, + 0.009964999742805958, + 0.014882789924740791, + 0.007909797132015228, + 0.01679813675582409, + 0.010006485506892204, + 0.010885138995945454, + 0.00894203968346119, + 0.030164970085024834, + 0.005554860457777977, + 0.010163442231714725, + 0.012017362751066685, + 0.02026936411857605, + 0.0054244184866547585, + 0.006889824755489826, + 0.005081638228148222, + 0.008032388053834438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.018183089792728424 + } + ], + "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/0000131/group_14/index_tensor([193, 193])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_14/index_tensor([193, 193])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ca0f9d719634598f661809eed227187430086394 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_14/index_tensor([193, 193])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In this shot, styled like high-speed sports photography, a dynamic, high-speed tracking shot maintains a tight, focused composition of Tom, an athlete performing a breathtaking, low-altitude wingsuit flight through a narrow canyon. He wears a stylish, aerodynamic wingsuit made of vibrant crimson and matte black technical fabric, perfectly fitting his muscular frame. A gleaming white full-face helmet and dark mirrored visor obscure his expression, reflecting the twisting, turbulent canyon walls. As he banks sharply to the left with decisive, controlled body movements, he comes incredibly close to the jagged, rust-colored rock face, stirring up small plumes of dust from the air currents. The camera rolls with him, capturing the immense force of his suit's wing fabric as it ripples and vibrates violently at sheer speed, a visual testament to his speed. The midday sun glints off his visor, casting deep, fast-moving shadows across the tiered rock, while the background fades into a blur of ochres and browns, contrasting sharply with the bright blue sky visible above. The entire scene is captured with the crisp, frozen-in-a-moment clarity of a high shutter speed photograph, conveying an intense atmosphere of pure, adrenaline-driven focus and high-stakes athleticism.", + "seed": 25104, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_14/index_193_seed_25104_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": -0.10249765496701002, + "vlm_logprob": -0.10249765496701002, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.10249765496701002, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3719654381275177 + ], + "opposite_logprob_tokens": [ + -1.2469654083251953 + ], + "logprob_reward": -0.3719654381275177, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3719654381275177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14369899034500122, + 0.19289818406105042, + 0.1440867781639099, + 0.06892930716276169, + 0.15185439586639404, + 0.0383467935025692, + 0.037872135639190674, + 0.04531329125165939, + 0.037794679403305054, + 0.13920539617538452 + ], + "credit_map": [ + 0.011956158094108105, + 0.008258551359176636, + 0.01416347548365593, + 0.015470169484615326, + 0.01450524851679802, + 0.009641049429774284, + 0.032433681190013885, + 0.03727065771818161, + 0.021163035184144974, + 0.020153308287262917, + 0.0320810005068779, + 0.028592662885785103, + 0.02698296494781971, + 0.02128671295940876, + 0.018679030239582062, + 0.023959463462233543, + 0.011568884365260601, + 0.009822417981922626, + 0.02622525580227375, + 0.029635045677423477, + 0.01424962654709816, + 0.021064436063170433, + 0.01725412905216217, + 0.014266989193856716, + 0.006975752301514149, + 0.004128136206418276, + 0.01733694225549698, + 0.010897922329604626, + 0.005459418520331383, + 0.007450649049133062, + 0.010383093729615211, + 0.0062974002212285995, + 0.011779550462961197, + 0.01772143319249153, + 0.014120064675807953, + 0.018360158428549767, + 0.02507147379219532, + 0.028653360903263092, + 0.015473025850951672, + 0.02067534066736698, + 0.003472953801974654, + 0.004463547375053167, + 0.003911664709448814, + 0.004849625285714865, + 0.003475635778158903, + 0.008340854197740555, + 0.0042612263932824135, + 0.005571288056671619, + 0.0032638206612318754, + 0.0036877337843179703, + 0.004288157448172569, + 0.004139475990086794, + 0.005167503375560045, + 0.006782613229006529, + 0.005169794894754887, + 0.0053730374202132225, + 0.00455919373780489, + 0.0039746747352182865, + 0.00870880950242281, + 0.0039195716381073, + 0.004689110908657312, + 0.009467811323702335, + 0.004557232838124037, + 0.005436890758574009, + 0.003981791436672211, + 0.00540936691686511, + 0.00468461075797677, + 0.0044815437868237495, + 0.004680756013840437, + 0.006803708150982857, + 0.003763943212106824, + 0.003988957963883877, + 0.012982483021914959, + 0.02549351193010807, + 0.015333911404013634, + 0.020345749333500862, + 0.020065752789378166, + 0.012372438795864582, + 0.015153651125729084, + 0.01745789870619774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.3719654381275177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221508987247944 + ], + "opposite_logprob_tokens": [ + -4.264221668243408 + ], + "logprob_reward": -0.014221508987247944, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221508987247944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12733426690101624, + 0.13286122679710388, + 0.10262449830770493, + 0.11916419863700867, + 0.055171992629766464, + 0.0649271160364151, + 0.15408271551132202, + 0.09722544997930527, + 0.06596305966377258, + 0.08064550906419754 + ], + "credit_map": [ + 0.010148214176297188, + 0.012364218011498451, + 0.018242862075567245, + 0.018900198861956596, + 0.015374056063592434, + 0.01163818035274744, + 0.024210767820477486, + 0.01645577698945999, + 0.014330662786960602, + 0.012680082581937313, + 0.026443077251315117, + 0.02171066775918007, + 0.011523709632456303, + 0.013465932570397854, + 0.015547885559499264, + 0.017159203067421913, + 0.011697997339069843, + 0.009792058728635311, + 0.018294334411621094, + 0.019269036129117012, + 0.012254935689270496, + 0.008736810646951199, + 0.011838309466838837, + 0.010741010308265686, + 0.010182803496718407, + 0.009122028015553951, + 0.02633041888475418, + 0.01989073120057583, + 0.013542532920837402, + 0.014375907368957996, + 0.01471419632434845, + 0.011005579493939877, + 0.004087714944034815, + 0.008301828987896442, + 0.0064124842174351215, + 0.006399550009518862, + 0.009646548889577389, + 0.007299251388758421, + 0.006007087882608175, + 0.007017530035227537, + 0.008780119940638542, + 0.010797080583870411, + 0.0059783682227134705, + 0.006761185824871063, + 0.005767112597823143, + 0.011544082313776016, + 0.005747710354626179, + 0.009551456198096275, + 0.01669350266456604, + 0.018988903611898422, + 0.027018386870622635, + 0.013733009807765484, + 0.017351245507597923, + 0.02783019468188286, + 0.013689929619431496, + 0.01877754181623459, + 0.007615676149725914, + 0.008228031918406487, + 0.02587420679628849, + 0.00849121157079935, + 0.011279263533651829, + 0.014698033221065998, + 0.009275955148041248, + 0.011763066053390503, + 0.007086650468409061, + 0.007486296351999044, + 0.007192372810095549, + 0.008373155258595943, + 0.009678425267338753, + 0.007965079508721828, + 0.010082665830850601, + 0.008098416030406952, + 0.006504278630018234, + 0.009561256505548954, + 0.007131112739443779, + 0.015622977167367935, + 0.012877654284238815, + 0.009314813651144505, + 0.00969585683196783, + 0.009937558323144913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.014221508987247944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013383012264966965 + ], + "opposite_logprob_tokens": [ + -6.626338481903076 + ], + "logprob_reward": -0.0013383012264966965, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013383012264966965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08572907000780106, + 0.07315409183502197, + 0.06506099551916122, + 0.05561106279492378, + 0.07864120602607727, + 0.046765442937612534, + 0.520967960357666, + 0.025320542976260185, + 0.016438942402601242, + 0.03231066092848778 + ], + "credit_map": [ + 0.014122476801276207, + 0.005048753228038549, + 0.00891483761370182, + 0.01217757910490036, + 0.007411102764308453, + 0.00732358917593956, + 0.015447819605469704, + 0.015282907523214817, + 0.006932196673005819, + 0.007382302079349756, + 0.01431373879313469, + 0.012105227448046207, + 0.008247456513345242, + 0.007586683612316847, + 0.0075132218189537525, + 0.00907326489686966, + 0.0065191881731152534, + 0.0061646453104913235, + 0.013894992880523205, + 0.012811223044991493, + 0.006739139091223478, + 0.0045142085291445255, + 0.007359989453107119, + 0.007057614624500275, + 0.004725465551018715, + 0.003082599490880966, + 0.014043314382433891, + 0.010544843040406704, + 0.0042739869095385075, + 0.007651443127542734, + 0.006253034342080355, + 0.00503637595102191, + 0.00858408585190773, + 0.010680940933525562, + 0.008873102255165577, + 0.01134492177516222, + 0.008544571697711945, + 0.012397531419992447, + 0.008689170703291893, + 0.009526879526674747, + 0.004325279965996742, + 0.0031779289711266756, + 0.005428697448223829, + 0.004524331074208021, + 0.003655982669442892, + 0.012800454162061214, + 0.006106951739639044, + 0.006745817139744759, + 0.05339872092008591, + 0.06483545899391174, + 0.0795135647058487, + 0.04847380146384239, + 0.07039730250835419, + 0.08851663023233414, + 0.053246282041072845, + 0.06258615851402283, + 0.0017228856449946761, + 0.001967406366020441, + 0.008736266754567623, + 0.002298749517649412, + 0.0016656029038131237, + 0.0037752806674689054, + 0.00170042528770864, + 0.003453926183283329, + 0.0017380941426381469, + 0.0018618000904098153, + 0.002067956607788801, + 0.0019025685032829642, + 0.0023295313585549593, + 0.002628740854561329, + 0.0020942389965057373, + 0.0018160110339522362, + 0.0031231932807713747, + 0.005752956494688988, + 0.004119975958019495, + 0.004054611083120108, + 0.0039359647780656815, + 0.004106000065803528, + 0.0031608629506081343, + 0.004057095851749182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0013383012264966965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002159208059311 + ], + "opposite_logprob_tokens": [ + -2.695021629333496 + ], + "logprob_reward": -0.07002159208059311, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002159208059311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24602124094963074, + 0.1038324311375618, + 0.12247428297996521, + 0.05052635446190834, + 0.10404305905103683, + 0.08933745324611664, + 0.10417277365922928, + 0.06625598669052124, + 0.06264431774616241, + 0.0506921224296093 + ], + "credit_map": [ + 0.03489977866411209, + 0.013991232961416245, + 0.01759924180805683, + 0.019356591627001762, + 0.01585269905626774, + 0.013810936361551285, + 0.08103007078170776, + 0.049480706453323364, + 0.015048336237668991, + 0.011830716393887997, + 0.021275628358125687, + 0.010638109408318996, + 0.009144133888185024, + 0.008971765637397766, + 0.014788015745580196, + 0.012135733850300312, + 0.013631056994199753, + 0.012740897946059704, + 0.019340218976140022, + 0.01912657357752323, + 0.01167734619230032, + 0.016455549746751785, + 0.01692917011678219, + 0.012573467567563057, + 0.0038792788982391357, + 0.0032710230443626642, + 0.006391813047230244, + 0.008424922823905945, + 0.004814251326024532, + 0.012438165955245495, + 0.0041287485510110855, + 0.007178151980042458, + 0.00791605468839407, + 0.013824890367686749, + 0.009640681557357311, + 0.012335886247456074, + 0.01772964373230934, + 0.017789281904697418, + 0.011052030138671398, + 0.013754589483141899, + 0.008174769580364227, + 0.007531957235187292, + 0.013156713917851448, + 0.009179842658340931, + 0.006036414299160242, + 0.023495804518461227, + 0.00875723548233509, + 0.013004723004996777, + 0.010931640863418579, + 0.008309934288263321, + 0.018777498975396156, + 0.010467620566487312, + 0.00857359729707241, + 0.022263694554567337, + 0.010403978638350964, + 0.014444807544350624, + 0.004785597790032625, + 0.005121678579598665, + 0.016680749133229256, + 0.004903287161141634, + 0.008510924875736237, + 0.01384749822318554, + 0.003888282924890518, + 0.008517967537045479, + 0.007383189629763365, + 0.007641037926077843, + 0.007328823674470186, + 0.006830458529293537, + 0.008852279745042324, + 0.010131402872502804, + 0.007157057989388704, + 0.007320065051317215, + 0.004629637114703655, + 0.005823346786201, + 0.008049668744206429, + 0.011820645071566105, + 0.004880480468273163, + 0.00354079925455153, + 0.0036969222128391266, + 0.008250623941421509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.07002159208059311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941434413194656 + ], + "opposite_logprob_tokens": [ + -2.929941415786743 + ], + "logprob_reward": -0.054941434413194656, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941434413194656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2628595232963562, + 0.09243256598711014, + 0.19733120501041412, + 0.06897038966417313, + 0.05923039838671684, + 0.04038815572857857, + 0.09928218275308609, + 0.04822646081447601, + 0.06906895339488983, + 0.06221015006303787 + ], + "credit_map": [ + 0.02507208101451397, + 0.01613701321184635, + 0.02659170888364315, + 0.021679837256669998, + 0.02027038112282753, + 0.015116263180971146, + 0.05837082862854004, + 0.07962141931056976, + 0.011760839261114597, + 0.008029979653656483, + 0.01859201490879059, + 0.012050065211951733, + 0.009345236234366894, + 0.007005606312304735, + 0.012914574705064297, + 0.012734245508909225, + 0.022022904828190804, + 0.01676749624311924, + 0.028729358687996864, + 0.03961670398712158, + 0.01702488772571087, + 0.02793494611978531, + 0.02120121754705906, + 0.0240336824208498, + 0.005282455123960972, + 0.004500970710068941, + 0.015958938747644424, + 0.01383821852505207, + 0.004981196951121092, + 0.009685427881777287, + 0.008018597029149532, + 0.0067045725882053375, + 0.004689085762947798, + 0.005743399728089571, + 0.0060681444592773914, + 0.0068888976238667965, + 0.006228829268366098, + 0.014163993299007416, + 0.006203381344676018, + 0.009244665503501892, + 0.003777379635721445, + 0.005227995570749044, + 0.0043519241735339165, + 0.005289142020046711, + 0.0038151403423398733, + 0.008491423912346363, + 0.004340468440204859, + 0.005094681400805712, + 0.008068561553955078, + 0.011593256145715714, + 0.013026447966694832, + 0.01110144518315792, + 0.00864222552627325, + 0.022663148120045662, + 0.009852724149823189, + 0.014334364794194698, + 0.004300456494092941, + 0.0035743550397455692, + 0.010015513747930527, + 0.003966262564063072, + 0.005231594666838646, + 0.010285898111760616, + 0.0037165358662605286, + 0.0071358392015099525, + 0.008636494167149067, + 0.011678623035550117, + 0.010913826525211334, + 0.009380683302879333, + 0.007839195430278778, + 0.0061119720339775085, + 0.007292204070836306, + 0.007215958088636398, + 0.0074378191493451595, + 0.00863618217408657, + 0.008323662914335728, + 0.006883122026920319, + 0.009953542612493038, + 0.006052753422409296, + 0.0076040588319301605, + 0.007319002877920866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.054941434413194656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In this shot, styled like high-speed sports photography, a dynamic, high-speed tracking shot maintains a tight, focused composition of Tom, an athlete performing a breathtaking, low-altitude wingsuit flight through a narrow canyon. He wears a stylish, aerodynamic wingsuit made of vibrant crimson and matte black technical fabric, perfectly fitting his muscular frame. A gleaming white full-face helmet and dark mirrored visor obscure his expression, reflecting the twisting, turbulent canyon walls. As he banks sharply to the left with decisive, controlled body movements, he comes incredibly close to the jagged, rust-colored rock face, stirring up small plumes of dust from the air currents. The camera rolls with him, capturing the immense force of his suit's wing fabric as it ripples and vibrates violently at sheer speed, a visual testament to his speed. The midday sun glints off his visor, casting deep, fast-moving shadows across the tiered rock, while the background fades into a blur of ochres and browns, contrasting sharply with the bright blue sky visible above. The entire scene is captured with the crisp, frozen-in-a-moment clarity of a high shutter speed photograph, conveying an intense atmosphere of pure, adrenaline-driven focus and high-stakes athleticism.", + "seed": 25105, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_14/index_193_seed_25105_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": -0.1701596020720899, + "vlm_logprob": -0.1701596020720899, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.1701596020720899, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6777083873748779 + ], + "opposite_logprob_tokens": [ + -0.8027083873748779 + ], + "logprob_reward": -0.6777083873748779, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6777083873748779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12457321584224701, + 0.1343526542186737, + 0.12191085517406464, + 0.07926169037818909, + 0.13329550623893738, + 0.10205680876970291, + 0.12293273210525513, + 0.0737607479095459, + 0.0544123537838459, + 0.053443457931280136 + ], + "credit_map": [ + 0.0213735643774271, + 0.011155321262776852, + 0.01992099918425083, + 0.016644610092043877, + 0.010104101151227951, + 0.007640509866178036, + 0.015845181420445442, + 0.021888932213187218, + 0.017842518165707588, + 0.019926298409700394, + 0.020121948793530464, + 0.015095893293619156, + 0.015383635647594929, + 0.013395071029663086, + 0.012003343552350998, + 0.02058393508195877, + 0.009009810164570808, + 0.013965566642582417, + 0.015773389488458633, + 0.025566738098859787, + 0.014858502894639969, + 0.011510159820318222, + 0.013970200903713703, + 0.01725647784769535, + 0.007880955003201962, + 0.008961196057498455, + 0.009643585421144962, + 0.012080018408596516, + 0.009455887600779533, + 0.009810325689613819, + 0.010835173539817333, + 0.010594542138278484, + 0.01668298803269863, + 0.010679924860596657, + 0.022129932418465614, + 0.012271560728549957, + 0.0259614996612072, + 0.012397391721606255, + 0.014275122433900833, + 0.018897071480751038, + 0.01628875359892845, + 0.016697945073246956, + 0.01495302002876997, + 0.010928862728178501, + 0.009231358766555786, + 0.010188818909227848, + 0.011763041839003563, + 0.012004997581243515, + 0.02254112809896469, + 0.012193860486149788, + 0.025405386462807655, + 0.0209621824324131, + 0.010914722457528114, + 0.010449245572090149, + 0.00995505042374134, + 0.010511145927011967, + 0.010452432557940483, + 0.007101629860699177, + 0.010851315222680569, + 0.00807847548276186, + 0.01187038328498602, + 0.006383873522281647, + 0.0091991713270545, + 0.009823461063206196, + 0.013884886167943478, + 0.00722248200327158, + 0.006760275922715664, + 0.00495125912129879, + 0.004739420488476753, + 0.006394402123987675, + 0.004728552885353565, + 0.0057310676202178, + 0.0065942066721618176, + 0.005438621621578932, + 0.00827658362686634, + 0.0063256737776100636, + 0.007720086257904768, + 0.005858155433088541, + 0.007051692344248295, + 0.006178433075547218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.6777083873748779 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012575025670230389 + ], + "opposite_logprob_tokens": [ + -4.387575149536133 + ], + "logprob_reward": -0.012575025670230389, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012575025670230389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22396525740623474, + 0.17173147201538086, + 0.05627566576004028, + 0.04632191359996796, + 0.12818622589111328, + 0.07671502232551575, + 0.08022340387105942, + 0.0864572823047638, + 0.0766463503241539, + 0.05347738787531853 + ], + "credit_map": [ + 0.052829448133707047, + 0.0220051109790802, + 0.024256184697151184, + 0.024660050868988037, + 0.021135615184903145, + 0.020962681621313095, + 0.023405008018016815, + 0.03471115976572037, + 0.027778904885053635, + 0.022832123562693596, + 0.026753831654787064, + 0.02275712601840496, + 0.020731240510940552, + 0.011714586056768894, + 0.01657927595078945, + 0.02258438803255558, + 0.004836967680603266, + 0.0074270498007535934, + 0.00593454809859395, + 0.009521950967609882, + 0.0069105797447264194, + 0.0068942224606871605, + 0.006406539585441351, + 0.008343808352947235, + 0.004415340721607208, + 0.005556874442845583, + 0.005180772393941879, + 0.005260147154331207, + 0.007756528910249472, + 0.006098183337599039, + 0.005328214727342129, + 0.006725853309035301, + 0.02187628671526909, + 0.010466137900948524, + 0.020984215661883354, + 0.011649628169834614, + 0.018166711553931236, + 0.01587381772696972, + 0.012384274043142796, + 0.0167851522564888, + 0.011373444460332394, + 0.00688820518553257, + 0.015776580199599266, + 0.011596085503697395, + 0.007484845817089081, + 0.006914312485605478, + 0.006899446249008179, + 0.0097820945084095, + 0.008235987275838852, + 0.014196553267538548, + 0.009623351506888866, + 0.00955055933445692, + 0.009078223258256912, + 0.010450808331370354, + 0.009322697296738625, + 0.00976521335542202, + 0.008126797154545784, + 0.010328102856874466, + 0.013014203868806362, + 0.01089421845972538, + 0.01597326621413231, + 0.0065823933109641075, + 0.011306012980639935, + 0.010232283733785152, + 0.014284725300967693, + 0.008201110176742077, + 0.007668233476579189, + 0.006084350403398275, + 0.008887730538845062, + 0.013085421174764633, + 0.008830761536955833, + 0.009604014456272125, + 0.007806684821844101, + 0.005499975755810738, + 0.00737642589956522, + 0.007825393229722977, + 0.009920291602611542, + 0.0049026114866137505, + 0.0056308964267373085, + 0.004515110049396753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.012575025670230389 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257826764136553 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257826764136553, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257826764136553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24759355187416077, + 0.2960362136363983, + 0.03919893130660057, + 0.011620134115219116, + 0.039200522005558014, + 0.07763572037220001, + 0.025675445795059204, + 0.03149344027042389, + 0.20672118663787842, + 0.02482483722269535 + ], + "credit_map": [ + 0.03719077631831169, + 0.013794046826660633, + 0.055257126688957214, + 0.01660020276904106, + 0.014185329899191856, + 0.02453864924609661, + 0.0244857557117939, + 0.06154169514775276, + 0.0316157266497612, + 0.05378882959485054, + 0.05910971760749817, + 0.02816990204155445, + 0.03283986449241638, + 0.025518344715237617, + 0.019170785322785378, + 0.045823074877262115, + 0.0038297201972454786, + 0.004778077360242605, + 0.004723073914647102, + 0.006976172793656588, + 0.004331185016781092, + 0.004058170132339001, + 0.004327984992414713, + 0.0061745536513626575, + 0.0013292841613292694, + 0.0012764493003487587, + 0.0008505460573360324, + 0.0013905842788517475, + 0.0018535391427576542, + 0.0016984902322292328, + 0.0015683530364185572, + 0.001652889302931726, + 0.0062734452076256275, + 0.003594404086470604, + 0.006644824985414743, + 0.003635843051597476, + 0.004539420362561941, + 0.00467006815597415, + 0.004175505600869656, + 0.00566701078787446, + 0.012578090652823448, + 0.010518758557736874, + 0.01171176042407751, + 0.007849709130823612, + 0.008580796420574188, + 0.007924307137727737, + 0.006990897934883833, + 0.01148140523582697, + 0.004866429138928652, + 0.003879417898133397, + 0.0031814510002732277, + 0.0028915260918438435, + 0.002752334577962756, + 0.0031796484254300594, + 0.002724139951169491, + 0.002200502436608076, + 0.0033458969555795193, + 0.003933533560484648, + 0.004958341829478741, + 0.004081166815012693, + 0.003995927982032299, + 0.002130000852048397, + 0.004398284014314413, + 0.004650291986763477, + 0.048081520944833755, + 0.025016458705067635, + 0.022173523902893066, + 0.01653393730521202, + 0.021043600514531136, + 0.03255122900009155, + 0.015702413395047188, + 0.025618532672524452, + 0.0026918291114270687, + 0.0026963152922689915, + 0.004459458868950605, + 0.0028329254128038883, + 0.003410318633541465, + 0.0031770796049386263, + 0.0024339789524674416, + 0.0031229332089424133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.005257826764136553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495971441268921 + ], + "opposite_logprob_tokens": [ + -2.929959774017334 + ], + "logprob_reward": -0.05495971441268921, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495971441268921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19858767092227936, + 0.06700879335403442, + 0.0882163792848587, + 0.04223167523741722, + 0.20722153782844543, + 0.16959306597709656, + 0.05527721345424652, + 0.10580261796712875, + 0.03853233531117439, + 0.027528773993253708 + ], + "credit_map": [ + 0.023456061258912086, + 0.013355281203985214, + 0.04122082516551018, + 0.0263780876994133, + 0.017107928171753883, + 0.01459493301808834, + 0.02792871743440628, + 0.03454580903053284, + 0.008239181712269783, + 0.007740509230643511, + 0.011811316013336182, + 0.010231198742985725, + 0.00823394488543272, + 0.006483626086264849, + 0.005790173541754484, + 0.00847883615642786, + 0.016301363706588745, + 0.008101110346615314, + 0.01178858894854784, + 0.010047473013401031, + 0.009167498908936977, + 0.009922442026436329, + 0.009822600521147251, + 0.013065294362604618, + 0.004941453691571951, + 0.005365828983485699, + 0.004279688466340303, + 0.005273537710309029, + 0.0051873233169317245, + 0.005433785263448954, + 0.005450352095067501, + 0.006299699191004038, + 0.023831164464354515, + 0.02494145557284355, + 0.025845257565379143, + 0.01997670717537403, + 0.02294992282986641, + 0.042057644575834274, + 0.0209074504673481, + 0.02671191096305847, + 0.0248245932161808, + 0.014820345677435398, + 0.038028888404369354, + 0.02579321525990963, + 0.013791072182357311, + 0.012435626238584518, + 0.016920005902647972, + 0.02297930046916008, + 0.009367293678224087, + 0.0041921003721654415, + 0.011297820135951042, + 0.010013610124588013, + 0.0036075955722481012, + 0.006364255212247372, + 0.0044512031599879265, + 0.00598332891240716, + 0.008949631825089455, + 0.013811028562486172, + 0.014983891509473324, + 0.014611995778977871, + 0.016785452142357826, + 0.008032702840864658, + 0.014511195942759514, + 0.014116708189249039, + 0.0080949068069458, + 0.004383890423923731, + 0.004775692708790302, + 0.003575972281396389, + 0.004411403089761734, + 0.004268189426511526, + 0.005583575461059809, + 0.0034387020859867334, + 0.004853864200413227, + 0.002417507115751505, + 0.004866695497184992, + 0.0028069016989320517, + 0.0031576587352901697, + 0.0033608563244342804, + 0.0023202472366392612, + 0.0037450408563017845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.05495971441268921 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029705613851547 + ], + "opposite_logprob_tokens": [ + -2.350296974182129 + ], + "logprob_reward": -0.10029705613851547, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029705613851547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12858439981937408, + 0.1109204888343811, + 0.041517022997140884, + 0.036662738770246506, + 0.3809368908405304, + 0.12774355709552765, + 0.06274538487195969, + 0.03840560466051102, + 0.03276810795068741, + 0.03971586376428604 + ], + "credit_map": [ + 0.02589774690568447, + 0.014592086896300316, + 0.015231735073029995, + 0.017066190019249916, + 0.014276153407990932, + 0.009920762851834297, + 0.01334347389638424, + 0.018256252631545067, + 0.012367571704089642, + 0.01651730388402939, + 0.023754607886075974, + 0.010321876034140587, + 0.013748902827501297, + 0.010078279301524162, + 0.007748282980173826, + 0.016383660957217216, + 0.004382629878818989, + 0.005292810499668121, + 0.004804519005119801, + 0.008154564537107944, + 0.004762273747473955, + 0.004063865169882774, + 0.0043512992560863495, + 0.005705057643353939, + 0.004535936750471592, + 0.0032412949949502945, + 0.005255304276943207, + 0.004314754623919725, + 0.006098321173340082, + 0.003746557980775833, + 0.004093176685273647, + 0.005377393215894699, + 0.05129238963127136, + 0.031165748834609985, + 0.06028198450803757, + 0.03323395922780037, + 0.05961428955197334, + 0.048862915486097336, + 0.041492730379104614, + 0.05499286204576492, + 0.02065952867269516, + 0.019056716933846474, + 0.021659787744283676, + 0.012818418443202972, + 0.012331055477261543, + 0.012138777412474155, + 0.01356132049113512, + 0.015517954714596272, + 0.012812820263206959, + 0.0033982452005147934, + 0.012452648021280766, + 0.007861132733523846, + 0.004472821950912476, + 0.004803542047739029, + 0.005006522871553898, + 0.011937648057937622, + 0.004992838483303785, + 0.006205975543707609, + 0.005794037599116564, + 0.004751450847834349, + 0.003573511028662324, + 0.0030238116160035133, + 0.004872768186032772, + 0.005191212985664606, + 0.004851729143410921, + 0.003078416921198368, + 0.00421222485601902, + 0.0030048161279410124, + 0.0034733915235847235, + 0.005964442156255245, + 0.0033175144344568253, + 0.004865575581789017, + 0.003353847423568368, + 0.004291154444217682, + 0.006807052996009588, + 0.004743259400129318, + 0.0070614079013466835, + 0.00430254777893424, + 0.003748690476641059, + 0.005407905671745539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.10029705613851547 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In this shot, styled like high-speed sports photography, a dynamic, high-speed tracking shot maintains a tight, focused composition of Tom, an athlete performing a breathtaking, low-altitude wingsuit flight through a narrow canyon. He wears a stylish, aerodynamic wingsuit made of vibrant crimson and matte black technical fabric, perfectly fitting his muscular frame. A gleaming white full-face helmet and dark mirrored visor obscure his expression, reflecting the twisting, turbulent canyon walls. As he banks sharply to the left with decisive, controlled body movements, he comes incredibly close to the jagged, rust-colored rock face, stirring up small plumes of dust from the air currents. The camera rolls with him, capturing the immense force of his suit's wing fabric as it ripples and vibrates violently at sheer speed, a visual testament to his speed. The midday sun glints off his visor, casting deep, fast-moving shadows across the tiered rock, while the background fades into a blur of ochres and browns, contrasting sharply with the bright blue sky visible above. The entire scene is captured with the crisp, frozen-in-a-moment clarity of a high shutter speed photograph, conveying an intense atmosphere of pure, adrenaline-driven focus and high-stakes athleticism.", + "seed": 25106, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_14/index_193_seed_25106_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": -0.10078190276399254, + "vlm_logprob": -0.10078190276399254, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.10078190276399254, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3743688762187958 + ], + "opposite_logprob_tokens": [ + -1.2493689060211182 + ], + "logprob_reward": -0.3743688762187958, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3743688762187958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041972722858190536, + 0.04767530411481857, + 0.01141142938286066, + 0.010978913865983486, + 0.009006243199110031, + 0.005330111365765333, + 0.03761580213904381, + 0.011365275830030441, + 0.02695569023489952, + 0.7976884841918945 + ], + "credit_map": [ + 0.0038458567578345537, + 0.002719264477491379, + 0.008578219451010227, + 0.0025112915318459272, + 0.0023284468334168196, + 0.0035105838906019926, + 0.0031826987396925688, + 0.01529636699706316, + 0.004588253330439329, + 0.006164636928588152, + 0.010320380330085754, + 0.004895043559372425, + 0.0033234411384910345, + 0.0032404165249317884, + 0.0056220353581011295, + 0.009521099738776684, + 0.0027409831527620554, + 0.0007213623612187803, + 0.0022530534770339727, + 0.0009447508491575718, + 0.000866201298777014, + 0.0012266351841390133, + 0.0008791279979050159, + 0.0017793158767744899, + 0.0010211628396064043, + 0.0009908664505928755, + 0.0020548857282847166, + 0.0009271419257856905, + 0.0017317794263362885, + 0.0012433930533006787, + 0.0009061270975507796, + 0.0021035587415099144, + 0.0013408620143309236, + 0.000987298204563558, + 0.0016784628387540579, + 0.0007182852132245898, + 0.0018747862195596099, + 0.001165782567113638, + 0.0006694558542221785, + 0.0005713115097023547, + 0.0011480612447485328, + 0.0008124646847136319, + 0.0008178632124327123, + 0.0007157236686907709, + 0.0007343837060034275, + 0.0003767594462260604, + 0.00039583860780112445, + 0.0003290174645371735, + 0.004325706046074629, + 0.004848917480558157, + 0.004627975635230541, + 0.002929066773504019, + 0.008987129665911198, + 0.003624908858910203, + 0.0033064018934965134, + 0.004965700209140778, + 0.0012781517580151558, + 0.0006722967955283821, + 0.003445597132667899, + 0.0006931367097422481, + 0.0004919011262245476, + 0.0005192679818719625, + 0.0006779536488465965, + 0.003586972365155816, + 0.003773193573579192, + 0.0037597077898681164, + 0.0035405310336500406, + 0.0038399535696953535, + 0.0027380811516195536, + 0.0026653388049453497, + 0.0027750853914767504, + 0.0038638010155409575, + 0.06210226193070412, + 0.024302149191498756, + 0.283444881439209, + 0.03367982804775238, + 0.009849593974649906, + 0.025961702689528465, + 0.023900272324681282, + 0.33444783091545105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.3743688762187958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221156015992165 + ], + "opposite_logprob_tokens": [ + -4.26422119140625 + ], + "logprob_reward": -0.014221156015992165, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221156015992165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15551957488059998, + 0.09684059768915176, + 0.06528272479772568, + 0.03630872815847397, + 0.0185370035469532, + 0.023077256977558136, + 0.16880157589912415, + 0.05854719877243042, + 0.06349202990531921, + 0.3135932683944702 + ], + "credit_map": [ + 0.01437403541058302, + 0.01013980433344841, + 0.030262012034654617, + 0.010854830965399742, + 0.008991958573460579, + 0.014833830296993256, + 0.011817260645329952, + 0.05424585938453674, + 0.009156113490462303, + 0.010209379717707634, + 0.019252726808190346, + 0.011713007465004921, + 0.008815374225378036, + 0.009536180645227432, + 0.011582886800169945, + 0.016574926674365997, + 0.007229477632790804, + 0.008095696568489075, + 0.011343774385750294, + 0.006613915786147118, + 0.007985218428075314, + 0.007453586906194687, + 0.00702719297260046, + 0.009533864445984364, + 0.004291339311748743, + 0.003605826757848263, + 0.007177502382546663, + 0.002254595747217536, + 0.0061947377398610115, + 0.004005941562354565, + 0.00299971760250628, + 0.005779069848358631, + 0.0027096287813037634, + 0.0021713862661272287, + 0.003293401561677456, + 0.0018180512124672532, + 0.0030292035080492496, + 0.0025277440436184406, + 0.0016423477791249752, + 0.001345242140814662, + 0.002837940352037549, + 0.003227503038942814, + 0.003523381659761071, + 0.00495989341288805, + 0.0031146774999797344, + 0.0018605589866638184, + 0.001870872569270432, + 0.0016824312042444944, + 0.017423490062355995, + 0.022862451151013374, + 0.01785183511674404, + 0.012161793187260628, + 0.041051410138607025, + 0.019705966114997864, + 0.0149863101541996, + 0.02275831811130047, + 0.006324257701635361, + 0.005160111002624035, + 0.013993225060403347, + 0.004961831029504538, + 0.005954726133495569, + 0.00387283181771636, + 0.004356237128376961, + 0.01392398588359356, + 0.01027414295822382, + 0.009862280450761318, + 0.009971799328923225, + 0.008923890069127083, + 0.00497791450470686, + 0.005369924008846283, + 0.0050517660565674305, + 0.009060317650437355, + 0.024790411815047264, + 0.011565606109797955, + 0.10895047336816788, + 0.013226451352238655, + 0.005561319645494223, + 0.011888525448739529, + 0.011079496704041958, + 0.12653101980686188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.014221156015992165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17632804811000824, + 0.15145091712474823, + 0.2178266942501068, + 0.04368782043457031, + 0.02731025032699108, + 0.05076819658279419, + 0.10732914507389069, + 0.030107488855719566, + 0.049214158207178116, + 0.14597730338573456 + ], + "credit_map": [ + 0.01869661547243595, + 0.01413293182849884, + 0.03311841934919357, + 0.015090416185557842, + 0.012742176651954651, + 0.019113237038254738, + 0.014341526664793491, + 0.049092721194028854, + 0.011507518589496613, + 0.01843845471739769, + 0.0259476937353611, + 0.016798080876469612, + 0.020466739311814308, + 0.015404436737298965, + 0.019429776817560196, + 0.023458199575543404, + 0.03556198999285698, + 0.01975129172205925, + 0.038720760494470596, + 0.02204160951077938, + 0.02745714597404003, + 0.01841215416789055, + 0.02193489670753479, + 0.03394682705402374, + 0.0056450883857905865, + 0.005661976523697376, + 0.007845853455364704, + 0.0038910044822841883, + 0.004950440023094416, + 0.0038077489007264376, + 0.0035481208469718695, + 0.008337581530213356, + 0.00589333288371563, + 0.003974318504333496, + 0.005237016826868057, + 0.002651976188644767, + 0.002423512516543269, + 0.0029664840549230576, + 0.002056413097307086, + 0.0021071943920105696, + 0.007089558057487011, + 0.009074615314602852, + 0.007724941708147526, + 0.010083389468491077, + 0.0058378358371555805, + 0.00398345896974206, + 0.003026077290996909, + 0.00394832156598568, + 0.012443071231245995, + 0.014356249012053013, + 0.01424078457057476, + 0.009439036250114441, + 0.023597167804837227, + 0.010101121850311756, + 0.00906011275947094, + 0.014091597869992256, + 0.006034205202013254, + 0.0027285716496407986, + 0.0083111971616745, + 0.0023595637176185846, + 0.0022572411689907312, + 0.0022519719786942005, + 0.0022714443039149046, + 0.003893291112035513, + 0.008213447406888008, + 0.007648950908333063, + 0.007242159452289343, + 0.006740938872098923, + 0.003863674122840166, + 0.00435596564784646, + 0.004184492398053408, + 0.006964526604861021, + 0.01625973917543888, + 0.010228270664811134, + 0.037838105112314224, + 0.010094505734741688, + 0.006887657102197409, + 0.007504008710384369, + 0.0066374437883496284, + 0.05052756518125534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00464016804471612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204111501574516 + ], + "opposite_logprob_tokens": [ + -2.8120410442352295 + ], + "logprob_reward": -0.06204111501574516, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204111501574516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06269820779561996, + 0.030611548572778702, + 0.03905508294701576, + 0.014960465021431446, + 0.008731558918952942, + 0.0154504319652915, + 0.08255186676979065, + 0.008302122354507446, + 0.014570496045053005, + 0.7230682373046875 + ], + "credit_map": [ + 0.005563400220125914, + 0.003378194523975253, + 0.012509538792073727, + 0.0043096584267914295, + 0.0038191473577171564, + 0.0058056749403476715, + 0.00504063256084919, + 0.022271953523159027, + 0.002505243057385087, + 0.0039973813109099865, + 0.005824845749884844, + 0.0037902104668319225, + 0.002477225149050355, + 0.002488522557541728, + 0.003808023640885949, + 0.00572008965536952, + 0.006237071007490158, + 0.004064250271767378, + 0.0066076102666556835, + 0.004243971314281225, + 0.003927633631974459, + 0.003900544485077262, + 0.004353107884526253, + 0.005720887333154678, + 0.0027254889719188213, + 0.001363786868751049, + 0.0025973219890147448, + 0.0012419087579473853, + 0.001490505994297564, + 0.0014122945722192526, + 0.0008887401781976223, + 0.0032404158264398575, + 0.0012548634549602866, + 0.001084446907043457, + 0.001602928969077766, + 0.0010170291643589735, + 0.0011851718882098794, + 0.0010692793875932693, + 0.000799902540165931, + 0.0007179349777288735, + 0.002502339193597436, + 0.0014818866038694978, + 0.002379063284024596, + 0.004361943807452917, + 0.0010761107550933957, + 0.001196464872919023, + 0.0012830236228182912, + 0.0011695976136252284, + 0.009480676613748074, + 0.011479387991130352, + 0.009627892635762691, + 0.006410491652786732, + 0.019190119579434395, + 0.008057162165641785, + 0.007234332617372274, + 0.01107178907841444, + 0.0009153622086159885, + 0.0012743811821565032, + 0.0017504951683804393, + 0.0009049677173607051, + 0.0005631057429127395, + 0.00038917851634323597, + 0.0006059610168449581, + 0.0018986695213243365, + 0.0026731030084192753, + 0.0018091050442308187, + 0.002042837208136916, + 0.0015647748950868845, + 0.0013006044318899512, + 0.001267469022423029, + 0.0014295160071924329, + 0.0024830843321979046, + 0.0567469447851181, + 0.022363316267728806, + 0.25648394227027893, + 0.030377503484487534, + 0.009159666486084461, + 0.02379564754664898, + 0.02181963622570038, + 0.3023214638233185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.06204111501574516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048638198524713516 + ], + "opposite_logprob_tokens": [ + -3.048638105392456 + ], + "logprob_reward": -0.048638198524713516, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048638198524713516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06898270547389984, + 0.05399809032678604, + 0.03467234596610069, + 0.014512273482978344, + 0.007144901901483536, + 0.0227900892496109, + 0.04300027713179588, + 0.00944437738507986, + 0.04078849405050278, + 0.7046664953231812 + ], + "credit_map": [ + 0.007381252944469452, + 0.00405368534848094, + 0.01334159355610609, + 0.004761508200317621, + 0.00291124964132905, + 0.006727010477334261, + 0.004943556617945433, + 0.02486284449696541, + 0.004332613665610552, + 0.00757084134966135, + 0.00924428179860115, + 0.005111434496939182, + 0.004610534757375717, + 0.006676516495645046, + 0.005431140307337046, + 0.01102072186768055, + 0.006319752428680658, + 0.002921655075624585, + 0.007582533173263073, + 0.0026434508617967367, + 0.0037365476600825787, + 0.00319861201569438, + 0.002749263308942318, + 0.00552052678540349, + 0.0020051361061632633, + 0.002243377035483718, + 0.002405407140031457, + 0.0014353820588439703, + 0.0021346493158489466, + 0.0013397351140156388, + 0.0009001143043860793, + 0.0020484705455601215, + 0.0010969656286761165, + 0.0006937693106010556, + 0.0016651980113238096, + 0.0006945570930838585, + 0.0009769536554813385, + 0.0008014027262106538, + 0.0006465865881182253, + 0.0005694687133654952, + 0.003704853355884552, + 0.0032178431283682585, + 0.004135474096983671, + 0.0028529781848192215, + 0.0034048801753669977, + 0.0018532238900661469, + 0.0019452834967523813, + 0.0016755509423092008, + 0.004621466621756554, + 0.004608261864632368, + 0.004653668496757746, + 0.004047449212521315, + 0.009870785288512707, + 0.005556577816605568, + 0.003918895032256842, + 0.005723167210817337, + 0.001401002169586718, + 0.0013068284606561065, + 0.0020098446402698755, + 0.001060528215020895, + 0.0008115175878629088, + 0.0007789791561663151, + 0.0007686772150918841, + 0.0013069986598566175, + 0.0061736986972391605, + 0.005407402757555246, + 0.0058182617649436, + 0.0055113728158175945, + 0.004213699139654636, + 0.0033190911635756493, + 0.003929059486836195, + 0.00641590217128396, + 0.05528552085161209, + 0.02200029417872429, + 0.25040969252586365, + 0.02942427434027195, + 0.009055429138243198, + 0.023136081174016, + 0.021551983430981636, + 0.2938031554222107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.048638198524713516 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In this shot, styled like high-speed sports photography, a dynamic, high-speed tracking shot maintains a tight, focused composition of Tom, an athlete performing a breathtaking, low-altitude wingsuit flight through a narrow canyon. He wears a stylish, aerodynamic wingsuit made of vibrant crimson and matte black technical fabric, perfectly fitting his muscular frame. A gleaming white full-face helmet and dark mirrored visor obscure his expression, reflecting the twisting, turbulent canyon walls. As he banks sharply to the left with decisive, controlled body movements, he comes incredibly close to the jagged, rust-colored rock face, stirring up small plumes of dust from the air currents. The camera rolls with him, capturing the immense force of his suit's wing fabric as it ripples and vibrates violently at sheer speed, a visual testament to his speed. The midday sun glints off his visor, casting deep, fast-moving shadows across the tiered rock, while the background fades into a blur of ochres and browns, contrasting sharply with the bright blue sky visible above. The entire scene is captured with the crisp, frozen-in-a-moment clarity of a high shutter speed photograph, conveying an intense atmosphere of pure, adrenaline-driven focus and high-stakes athleticism.", + "seed": 25107, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_14/index_193_seed_25107_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": -0.11736695636063814, + "vlm_logprob": -0.11736695636063814, + "vlm_yes_no_margin": 3.0249999880790712, + "vlm_token_credit": -0.11736695636063814, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.45827656984329224 + ], + "opposite_logprob_tokens": [ + -1.0832765102386475 + ], + "logprob_reward": -0.45827656984329224, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.45827656984329224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07533688843250275, + 0.05470302700996399, + 0.1932666301727295, + 0.07615926861763, + 0.12153450399637222, + 0.04461057111620903, + 0.0686517208814621, + 0.11155633628368378, + 0.14572516083717346, + 0.10845587402582169 + ], + "credit_map": [ + 0.00959171075373888, + 0.015515478327870369, + 0.009098322130739689, + 0.010006050579249859, + 0.0069490051828324795, + 0.008656531572341919, + 0.006590745877474546, + 0.008929045870900154, + 0.006842681672424078, + 0.013918985612690449, + 0.0067665730603039265, + 0.007832251489162445, + 0.005433477461338043, + 0.0046628620475530624, + 0.0049144853837788105, + 0.0043317098170518875, + 0.04498906433582306, + 0.010116476565599442, + 0.026882480829954147, + 0.05647493526339531, + 0.008723407983779907, + 0.013346466235816479, + 0.017722448334097862, + 0.015011347830295563, + 0.008047058247029781, + 0.007445893250405788, + 0.013562985695898533, + 0.009084573946893215, + 0.009258551523089409, + 0.009001841768622398, + 0.009966499172151089, + 0.009791872464120388, + 0.036105960607528687, + 0.006654961965978146, + 0.04915120080113411, + 0.008916324004530907, + 0.0032854105811566114, + 0.002883949549868703, + 0.008047929033637047, + 0.006488759070634842, + 0.007207774091511965, + 0.004333047196269035, + 0.010298539884388447, + 0.0065931291319429874, + 0.003495476208627224, + 0.002572299912571907, + 0.003206617198884487, + 0.006903687957674265, + 0.012477030046284199, + 0.008840017952024937, + 0.006164082791656256, + 0.0032361168414354324, + 0.009102917276322842, + 0.005354252178221941, + 0.006226835306733847, + 0.017250467091798782, + 0.01195053942501545, + 0.02997744269669056, + 0.011397226713597775, + 0.008410525508224964, + 0.009146980941295624, + 0.00909771304577589, + 0.0068161399103701115, + 0.024759763851761818, + 0.03998623415827751, + 0.05740103870630264, + 0.004468508064746857, + 0.006533948238939047, + 0.007107864134013653, + 0.007371722254902124, + 0.007607597392052412, + 0.01524825394153595, + 0.011611074209213257, + 0.014859725721180439, + 0.008298578672111034, + 0.015173062682151794, + 0.01311617624014616, + 0.020528512075543404, + 0.009697569534182549, + 0.015171180479228497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.45827656984329224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014227267354726791 + ], + "opposite_logprob_tokens": [ + -4.264227390289307 + ], + "logprob_reward": -0.014227267354726791, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014227267354726791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0526370108127594, + 0.0934641882777214, + 0.10967780649662018, + 0.0769527330994606, + 0.4040779769420624, + 0.043445467948913574, + 0.0127747543156147, + 0.04906995967030525, + 0.11889864504337311, + 0.039001453667879105 + ], + "credit_map": [ + 0.007432152051478624, + 0.007852623239159584, + 0.005679120775312185, + 0.005816393531858921, + 0.0055436328984797, + 0.005803974810987711, + 0.007066785823553801, + 0.007442324887961149, + 0.009013909846544266, + 0.007817687466740608, + 0.009749306365847588, + 0.006470970343798399, + 0.011413885280489922, + 0.01909455843269825, + 0.012932300567626953, + 0.01697157323360443, + 0.024136357009410858, + 0.006621051579713821, + 0.018961068242788315, + 0.016767069697380066, + 0.008018087595701218, + 0.010733315721154213, + 0.010629108175635338, + 0.013811754062771797, + 0.01098095066845417, + 0.007905721664428711, + 0.012642374262213707, + 0.012720467522740364, + 0.008230545558035374, + 0.00883182231336832, + 0.008001093752682209, + 0.007639758288860321, + 0.11682599037885666, + 0.016784541308879852, + 0.17991377413272858, + 0.020301364362239838, + 0.015266750007867813, + 0.007133908104151487, + 0.029344530776143074, + 0.018507109954953194, + 0.007971672341227531, + 0.0075533767230808735, + 0.011437335051596165, + 0.004683776292949915, + 0.0038388355169445276, + 0.0022478201426565647, + 0.003197802696377039, + 0.0025148484855890274, + 0.002061407547444105, + 0.0016164635308086872, + 0.0023707025684416294, + 0.0010272584622725844, + 0.0010107571724802256, + 0.0008349984418600798, + 0.0010140370577573776, + 0.002839129650965333, + 0.004977153614163399, + 0.016849951818585396, + 0.0038261215668171644, + 0.002506646793335676, + 0.0033389851450920105, + 0.0030290421564131975, + 0.002227600198239088, + 0.012314457446336746, + 0.04213440790772438, + 0.03512181341648102, + 0.007334033492952585, + 0.0063351779244840145, + 0.006203262601047754, + 0.005974112078547478, + 0.007787567563354969, + 0.008008266799151897, + 0.003498160745948553, + 0.005578789860010147, + 0.002786488737910986, + 0.00510723190382123, + 0.005424083676189184, + 0.0054618557915091515, + 0.005058327689766884, + 0.006086512468755245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.014227267354726791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201176352798939 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201176352798939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201176352798939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16634365916252136, + 0.1030130684375763, + 0.07877358794212341, + 0.15847060084342957, + 0.06934969872236252, + 0.044992830604314804, + 0.05473709851503372, + 0.15504132211208344, + 0.07609165459871292, + 0.09318644553422928 + ], + "credit_map": [ + 0.025229793041944504, + 0.020105790346860886, + 0.01979379914700985, + 0.027626067399978638, + 0.01630573719739914, + 0.01755404844880104, + 0.01784352958202362, + 0.021884895861148834, + 0.01539222989231348, + 0.01316136959940195, + 0.012204037047922611, + 0.012031393125653267, + 0.011574032716453075, + 0.014455368742346764, + 0.012593453750014305, + 0.01160118356347084, + 0.013710380531847477, + 0.006194625049829483, + 0.011658050119876862, + 0.02253720536828041, + 0.004875504411756992, + 0.006998243276029825, + 0.007692492567002773, + 0.00510708661749959, + 0.02104783244431019, + 0.015979841351509094, + 0.029499094933271408, + 0.022654034197330475, + 0.017948715016245842, + 0.016293946653604507, + 0.018832335248589516, + 0.01621481217443943, + 0.020897874608635902, + 0.0032959883101284504, + 0.028421195223927498, + 0.004113428760319948, + 0.0023461754899472, + 0.0016035338630899787, + 0.004507077392190695, + 0.004164426121860743, + 0.008104688487946987, + 0.007694272790104151, + 0.010293396189808846, + 0.005410441663116217, + 0.003763512708246708, + 0.0021809148602187634, + 0.00370249361731112, + 0.0038431105203926563, + 0.011596950702369213, + 0.004459487274289131, + 0.00523283751681447, + 0.0026766310911625624, + 0.00382803613319993, + 0.005299590993672609, + 0.004572262056171894, + 0.017071306705474854, + 0.02211282029747963, + 0.029169777408242226, + 0.0242896918207407, + 0.014577975496649742, + 0.010732886381447315, + 0.013956657610833645, + 0.00904786679893732, + 0.031153643503785133, + 0.024188870564103127, + 0.03060740791261196, + 0.003613609354943037, + 0.0028468237724155188, + 0.0037653702311217785, + 0.0026503263507038355, + 0.002845688257366419, + 0.005573561415076256, + 0.010534449480473995, + 0.012616956606507301, + 0.007062820252031088, + 0.011390508152544498, + 0.011398485861718655, + 0.01953309029340744, + 0.007393839303404093, + 0.013256299309432507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.014201176352798939 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381012037396431 + ], + "opposite_logprob_tokens": [ + -3.2881011962890625 + ], + "logprob_reward": -0.0381012037396431, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381012037396431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09575600922107697, + 0.09038156270980835, + 0.10249356925487518, + 0.06375156342983246, + 0.3091018497943878, + 0.06266999989748001, + 0.06552714854478836, + 0.06543737649917603, + 0.07001546770334244, + 0.07486533373594284 + ], + "credit_map": [ + 0.010796424001455307, + 0.011701054871082306, + 0.011306541040539742, + 0.012282478623092175, + 0.011693349108099937, + 0.009990698657929897, + 0.014381175860762596, + 0.013604296371340752, + 0.010389932431280613, + 0.009375111199915409, + 0.00981291476637125, + 0.009254015982151031, + 0.011626550927758217, + 0.01373083796352148, + 0.012289555743336678, + 0.013902650214731693, + 0.023643366992473602, + 0.0070787230506539345, + 0.014765688218176365, + 0.03014897182583809, + 0.004878669511526823, + 0.008285176008939743, + 0.00703834742307663, + 0.006654633209109306, + 0.008733413182199001, + 0.007496821228414774, + 0.010770834982395172, + 0.01106285396963358, + 0.0076354434713721275, + 0.005493243224918842, + 0.006181647069752216, + 0.006377314683049917, + 0.09213931858539581, + 0.012264877557754517, + 0.13845418393611908, + 0.014847218059003353, + 0.008308482356369495, + 0.0051271673291921616, + 0.022995799779891968, + 0.014964819885790348, + 0.011113984510302544, + 0.010961618274450302, + 0.011297573335468769, + 0.009921514429152012, + 0.0048102946020662785, + 0.003844707040116191, + 0.004825350362807512, + 0.005894962698221207, + 0.010957076214253902, + 0.00957126822322607, + 0.006068359594792128, + 0.003881310112774372, + 0.008109886199235916, + 0.005271471571177244, + 0.005541507620364428, + 0.016126278787851334, + 0.0058539146557450294, + 0.01640763320028782, + 0.007629838772118092, + 0.004833596758544445, + 0.004329141229391098, + 0.005899823736399412, + 0.0039766584523022175, + 0.016506770625710487, + 0.0228059533983469, + 0.01856101118028164, + 0.004362409934401512, + 0.003988652490079403, + 0.0039181895554065704, + 0.006658877711743116, + 0.003553135087713599, + 0.006167250219732523, + 0.008598343469202518, + 0.012687308713793755, + 0.006496939342468977, + 0.00972247589379549, + 0.010627120733261108, + 0.007403777912259102, + 0.01051547471433878, + 0.008813898079097271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0381012037396431 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 athlete named Tom performing a low-altitude wingsuit flight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wingsuit made of vibrant crimson and matte black technical fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a narrow canyon with jagged, rust-colored rock faces and a bright blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera roll with the subject while capturing the suit's fabric rippling and vibrating violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style characterized by the crisp, frozen-in-a-moment clarity of a high shutter speed photograph?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062028564512729645 + ], + "opposite_logprob_tokens": [ + -2.812028646469116 + ], + "logprob_reward": -0.062028564512729645, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062028564512729645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06045510992407799, + 0.05157927796244621, + 0.325514554977417, + 0.15649883449077606, + 0.06200074404478073, + 0.045908547937870026, + 0.08513370901346207, + 0.06049022451043129, + 0.08131834119558334, + 0.0711006298661232 + ], + "credit_map": [ + 0.00836880598217249, + 0.007576111704111099, + 0.00634929072111845, + 0.007455059792846441, + 0.007531594019383192, + 0.0063160997815430164, + 0.008713243529200554, + 0.008144911378622055, + 0.006774470675736666, + 0.007396860513836145, + 0.012264486402273178, + 0.006862180307507515, + 0.004245489835739136, + 0.0042861588299274445, + 0.005480795633047819, + 0.004268837161362171, + 0.07515385001897812, + 0.01827176660299301, + 0.04600391536951065, + 0.11252233386039734, + 0.013367925770580769, + 0.019115764647722244, + 0.021711641922593117, + 0.019367385655641556, + 0.02647894062101841, + 0.025179432705044746, + 0.029346901923418045, + 0.0357479453086853, + 0.011173938401043415, + 0.008365891873836517, + 0.01020914874970913, + 0.009996642358601093, + 0.01669386215507984, + 0.003125481540337205, + 0.02503788098692894, + 0.003289339365437627, + 0.0027490493375808, + 0.0021182808559387922, + 0.004857117775827646, + 0.00412973714992404, + 0.008786614052951336, + 0.004634396638721228, + 0.013496959581971169, + 0.004472843371331692, + 0.0045894719660282135, + 0.00253484770655632, + 0.0028909991960972548, + 0.004502414725720882, + 0.01785515435039997, + 0.009352147579193115, + 0.0068100327625870705, + 0.004458578769117594, + 0.007218983955681324, + 0.007720900233834982, + 0.007354927249252796, + 0.024362994357943535, + 0.005743596237152815, + 0.020187120884656906, + 0.004809988662600517, + 0.0030408743768930435, + 0.004737711511552334, + 0.004098519682884216, + 0.00255491747520864, + 0.01531749777495861, + 0.01488326396793127, + 0.027766050770878792, + 0.005154096987098455, + 0.00513220950961113, + 0.004364193417131901, + 0.005680916365236044, + 0.00548706715926528, + 0.01285055186599493, + 0.008399351499974728, + 0.007748620584607124, + 0.0064013646915555, + 0.009952060878276825, + 0.009151134639978409, + 0.010758798569440842, + 0.010046495124697685, + 0.008642809465527534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.062028564512729645 + } + ], + "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/0000131/group_15/index_tensor([5169, 5169])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_15/index_tensor([5169, 5169])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90f70a405da8fa2e33dae88f98ce4d4f639a2f65 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_15/index_tensor([5169, 5169])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot captured from a high-angle, slightly bird's-eye view from a drone, the camera slowly and steadily follows the subject in synchronized, circular motion. The image reveals a minimalist world composed of pure black, pure white, and countless shades of gray in between. A vast, mirror-like expanse of gray water reflects the leaden sky, the two colors blending seamlessly, with almost no demarcation. In the center of the water, a simple marker, made from a few weathered bamboo poles, stands alone. A fisherman steers a narrow, dark-black wooden boat, its simple lines stained with a subtle matte sheen from years of use. With his back to the camera, the fisherman, hunched over, wears a wide, traditional bamboo hat woven from withered bamboo strips and a loose, dark-gray cotton-linen jacket. He gripped a long wooden oar tightly in both hands, each stroke appearing unhurried and rhythmic. The oar's impact on the water only stirred faint, rapidly expanding white ripples on the mirror-like surface. The boat moved slowly, at an almost meditative pace, in a perfect circle, circling the bamboo pole mark. The even, soft, diffuse light outlined the contours of everything without creating noticeable shadows, enhancing the flatness and abstraction of the painting. The overall atmosphere was imbued with the poetic and Zen-like qualities of an Eastern ink painting. An ultimate sense of tranquility, solitude, and eternity enveloped the entire scene.", + "seed": 26503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_15/index_5169_seed_26503_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": -0.02389736855402589, + "vlm_logprob": -0.02389736855402589, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02389736855402589, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.024902142584323883 + ], + "opposite_logprob_tokens": [ + -5.02490234375 + ], + "logprob_reward": -0.024902142584323883, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.024902142584323883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7188796401023865, + 0.0335957370698452, + 0.051966868340969086, + 0.018396614119410515, + 0.033051516860723495, + 0.03989948332309723, + 0.021147212013602257, + 0.026592010632157326, + 0.03450458124279976, + 0.021966325119137764 + ], + "credit_map": [ + 0.25075268745422363, + 0.02619118057191372, + 0.3102998435497284, + 0.029256630688905716, + 0.00795553158968687, + 0.028530966490507126, + 0.02020381949841976, + 0.04568898677825928, + 0.0037173398304730654, + 0.00318119116127491, + 0.0057063414715230465, + 0.004768108017742634, + 0.0025413427501916885, + 0.004379465710371733, + 0.00391036830842495, + 0.005391582380980253, + 0.006525490898638964, + 0.004451279528439045, + 0.008210104890167713, + 0.0063667744398117065, + 0.0033588665537536144, + 0.00531765865162015, + 0.009176403284072876, + 0.008560295216739178, + 0.002768796170130372, + 0.0020345617085695267, + 0.0031697144731879234, + 0.002222793409600854, + 0.0015370892360806465, + 0.0020564934238791466, + 0.001763261971063912, + 0.0028439031448215246, + 0.004066109657287598, + 0.0039059568662196398, + 0.006967872381210327, + 0.0040366509929299355, + 0.002769136568531394, + 0.003008612198755145, + 0.004010080825537443, + 0.004287101794034243, + 0.004342990927398205, + 0.004989308770745993, + 0.006461045239120722, + 0.003986602649092674, + 0.002973153255879879, + 0.007592399138957262, + 0.005118940956890583, + 0.004435045178979635, + 0.003003030549734831, + 0.0015802321722730994, + 0.005669384263455868, + 0.0019203381380066276, + 0.0013552571181207895, + 0.00357148889452219, + 0.0015264517860487103, + 0.002521032001823187, + 0.0031137552578002214, + 0.0018821600824594498, + 0.00455700745806098, + 0.0031848703511059284, + 0.0031341444700956345, + 0.005376134067773819, + 0.002495354972779751, + 0.0028485856018960476, + 0.003468533279374242, + 0.0023435200564563274, + 0.005588205065578222, + 0.003042985685169697, + 0.0038595842197537422, + 0.004724940285086632, + 0.0038636941462755203, + 0.00761311873793602, + 0.004037033766508102, + 0.0018505938351154327, + 0.0036061976570636034, + 0.001813307753764093, + 0.002794093918055296, + 0.003193813841789961, + 0.0020661880262196064, + 0.002605097135528922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.024902142584323883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.023321963846683502 + ], + "opposite_logprob_tokens": [ + -3.7733218669891357 + ], + "logprob_reward": -0.023321963846683502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023321963846683502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7711743712425232, + 0.026270391419529915, + 0.017223438248038292, + 0.01977861113846302, + 0.029063262045383453, + 0.019692471250891685, + 0.021188514307141304, + 0.04127922281622887, + 0.025455834344029427, + 0.02887384407222271 + ], + "credit_map": [ + 0.27115532755851746, + 0.027216186746954918, + 0.33457064628601074, + 0.029997335746884346, + 0.008395080454647541, + 0.029584607109427452, + 0.021029770374298096, + 0.04922553896903992, + 0.003907929640263319, + 0.0022443407215178013, + 0.005038043018430471, + 0.002904505468904972, + 0.0016238626558333635, + 0.002574334153905511, + 0.0035413317382335663, + 0.004436046816408634, + 0.0033548236824572086, + 0.0014863304095342755, + 0.0033437018282711506, + 0.0016937555046752095, + 0.0009435522370040417, + 0.0013694283552467823, + 0.0021515158005058765, + 0.0028803329914808273, + 0.004334139637649059, + 0.0012107450747862458, + 0.006248493678867817, + 0.0013637178344652057, + 0.0010439932812005281, + 0.0016789451474323869, + 0.0011227168142795563, + 0.0027758628129959106, + 0.0035408460535109043, + 0.003509524278342724, + 0.006714902352541685, + 0.0037217484787106514, + 0.0021989750675857067, + 0.002465651137754321, + 0.0032952497713267803, + 0.003616368630900979, + 0.0022347362246364355, + 0.0010366786736994982, + 0.004686486907303333, + 0.0019831196404993534, + 0.0010995363118126988, + 0.004887699615210295, + 0.0017610267968848348, + 0.002003189641982317, + 0.0030765465926378965, + 0.001762591302394867, + 0.004776947200298309, + 0.0016321409493684769, + 0.0013897627359256148, + 0.004781131632626057, + 0.0013745815958827734, + 0.0023948149755597115, + 0.00406321045011282, + 0.0038590484764426947, + 0.008141275495290756, + 0.0047937482595443726, + 0.004525857977569103, + 0.007144032511860132, + 0.004674957599490881, + 0.004077098332345486, + 0.003073431085795164, + 0.002270094584673643, + 0.004482685588300228, + 0.0038057907950133085, + 0.003725927323102951, + 0.0030114511027932167, + 0.0023041460663080215, + 0.00278231012634933, + 0.002575110411271453, + 0.0023358811158686876, + 0.0050098174251616, + 0.003567345440387726, + 0.002742012497037649, + 0.003505096537992358, + 0.002970380475744605, + 0.006168204825371504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.023321963846683502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.020585810765624046 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585810765624046, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585810765624046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8650171756744385, + 0.008875791914761066, + 0.01762377843260765, + 0.035348135977983475, + 0.005642218980938196, + 0.009191419929265976, + 0.013026460073888302, + 0.0174863301217556, + 0.009416576474905014, + 0.018372103571891785 + ], + "credit_map": [ + 0.3064192533493042, + 0.029230505228042603, + 0.3771696984767914, + 0.03443874791264534, + 0.008496951311826706, + 0.03125708922743797, + 0.022754790261387825, + 0.055250026285648346, + 0.0010308794444426894, + 0.0008621570304967463, + 0.0012118665035814047, + 0.0012168798130005598, + 0.000914638745598495, + 0.0014472948387265205, + 0.0010166827123612165, + 0.001175392186269164, + 0.0030819319654256105, + 0.0013268371112644672, + 0.003570287022739649, + 0.001700813532806933, + 0.0010247098980471492, + 0.0014314426807686687, + 0.002152330707758665, + 0.003335424466058612, + 0.007115701679140329, + 0.0034561704378575087, + 0.009953247383236885, + 0.003891527187079191, + 0.0019312226213514805, + 0.002597590209916234, + 0.0030862586572766304, + 0.0033164136111736298, + 0.0008837391505949199, + 0.0006891831289976835, + 0.0015005075838416815, + 0.0004975033225491643, + 0.0003878098214045167, + 0.00044937050552107394, + 0.0006136338924989104, + 0.0006204710807651281, + 0.0008271164260804653, + 0.0007376373396255076, + 0.0018856569658964872, + 0.0014734548749402165, + 0.0005750569980591536, + 0.001174934091977775, + 0.0014448382426053286, + 0.0010727238841354847, + 0.0018181534251198173, + 0.0009133171406574547, + 0.0025421190075576305, + 0.0011533335782587528, + 0.0008356444886885583, + 0.0033042840659618378, + 0.001100800116546452, + 0.0013588066212832928, + 0.002070395275950432, + 0.0013900185003876686, + 0.003450623946264386, + 0.0017915911739692092, + 0.002143969526514411, + 0.0027384806890040636, + 0.002181350253522396, + 0.0017198983114212751, + 0.0010805956553667784, + 0.000812712125480175, + 0.0015648280968889594, + 0.0013265602756291628, + 0.0009633192094042897, + 0.0012428618501871824, + 0.0011613027891144156, + 0.0012643951922655106, + 0.0015988154336810112, + 0.0015659425407648087, + 0.00271894340403378, + 0.002391335554420948, + 0.001405599294230342, + 0.0020661544986069202, + 0.002093187067657709, + 0.004532123450189829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.020585810765624046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.0076352194882929325 + ], + "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": [ + 0.7951003313064575, + 0.027661748230457306, + 0.012920298613607883, + 0.03977984935045242, + 0.007718790788203478, + 0.01695089228451252, + 0.042016565799713135, + 0.02416268363595009, + 0.01821000501513481, + 0.015478824265301228 + ], + "credit_map": [ + 0.28078901767730713, + 0.027608947828412056, + 0.34762275218963623, + 0.03105315938591957, + 0.007886986248195171, + 0.028591549023985863, + 0.02115245908498764, + 0.050395429134368896, + 0.0034107135143131018, + 0.003156876191496849, + 0.004353420343250036, + 0.0035330462269484997, + 0.0017262218752875924, + 0.0032957694493234158, + 0.004106160718947649, + 0.004079537466168404, + 0.0015837501268833876, + 0.0009841641876846552, + 0.0024379633832722902, + 0.0014077880186960101, + 0.0006776745431125164, + 0.001337444526143372, + 0.0021585056092590094, + 0.0023330082185566425, + 0.008375328965485096, + 0.0032020017970353365, + 0.011198321357369423, + 0.004402862396091223, + 0.0019216204527765512, + 0.0034015323035418987, + 0.0031433352269232273, + 0.00413484638556838, + 0.0007976570050232112, + 0.0007045005331747234, + 0.0018222358776256442, + 0.0009747357107698917, + 0.0006388267502188683, + 0.0007991243037395179, + 0.0009808327304199338, + 0.0010008772369474173, + 0.0016417786246165633, + 0.001107870601117611, + 0.003377180313691497, + 0.0030230563133955, + 0.0011387785198166966, + 0.0023278736043721437, + 0.0022336062975227833, + 0.0021007456816732883, + 0.0038376045413315296, + 0.003937176428735256, + 0.00685504637658596, + 0.005328531377017498, + 0.00476966192945838, + 0.005517355166375637, + 0.006953890435397625, + 0.004817298147827387, + 0.0024684700183570385, + 0.002687403466552496, + 0.003663629060611129, + 0.002689545974135399, + 0.0032836319878697395, + 0.0035688800271600485, + 0.002934335730969906, + 0.002866785740479827, + 0.0019461234332993627, + 0.0016419884050264955, + 0.002515649888664484, + 0.0026254872791469097, + 0.003592969849705696, + 0.002637529978528619, + 0.0015002986183390021, + 0.0017499566311016679, + 0.0018163779750466347, + 0.0014685027999803424, + 0.0028850422240793705, + 0.0015132868429645896, + 0.0023543068673461676, + 0.0018520466983318329, + 0.0018449778435751796, + 0.0017442824319005013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.0076352194882929325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.04304170608520508 + ], + "opposite_logprob_tokens": [ + -3.168041706085205 + ], + "logprob_reward": -0.04304170608520508, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304170608520508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7027304172515869, + 0.01676098071038723, + 0.01566299982368946, + 0.0572325736284256, + 0.015637172386050224, + 0.01593703217804432, + 0.03045874461531639, + 0.09025274962186813, + 0.044894199818372726, + 0.010433214716613293 + ], + "credit_map": [ + 0.2480105757713318, + 0.024258848279714584, + 0.30705246329307556, + 0.027797695249319077, + 0.007085348945111036, + 0.025114115327596664, + 0.018704436719417572, + 0.044706907123327255, + 0.0036169919185340405, + 0.001985755516216159, + 0.004222647286951542, + 0.0013537792256101966, + 0.0008395404438488185, + 0.0018083298346027732, + 0.0012821359559893608, + 0.0016517979092895985, + 0.0014511195477098227, + 0.0017870506271719933, + 0.0023016042541712523, + 0.0016568211140111089, + 0.0012800234835594893, + 0.0017127094324678183, + 0.0024696053005754948, + 0.0030040647834539413, + 0.014038758352398872, + 0.003625396639108658, + 0.020613564178347588, + 0.004164539743214846, + 0.0019634789787232876, + 0.003513023955747485, + 0.0036537654232233763, + 0.0056600421667099, + 0.0025375126861035824, + 0.0014546557795256376, + 0.004107485059648752, + 0.0017491568578407168, + 0.0008004205883480608, + 0.0014491949696093798, + 0.0017170002683997154, + 0.0018217446049675345, + 0.0014737191377207637, + 0.0017076556105166674, + 0.002174583962187171, + 0.0026050391606986523, + 0.0015158425085246563, + 0.0022564218379557133, + 0.002265406306833029, + 0.0019383630715310574, + 0.0032789097167551517, + 0.002208235440775752, + 0.006409252993762493, + 0.002983088605105877, + 0.0026117421220988035, + 0.005884972866624594, + 0.003864865517243743, + 0.003217674558982253, + 0.011202404275536537, + 0.0065803201869130135, + 0.017773395404219627, + 0.00941808708012104, + 0.010272499173879623, + 0.016359103843569756, + 0.00975961796939373, + 0.008887318894267082, + 0.005010878201574087, + 0.0031280601397156715, + 0.008133409544825554, + 0.003029938554391265, + 0.005510701797902584, + 0.009577435441315174, + 0.0037264602724462748, + 0.006777310278266668, + 0.0011459378292784095, + 0.001066112658008933, + 0.0018334761261940002, + 0.001436143647879362, + 0.0009498207946307957, + 0.0012142742052674294, + 0.0010756428819149733, + 0.0017118058167397976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.04304170608520508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot captured from a high-angle, slightly bird's-eye view from a drone, the camera slowly and steadily follows the subject in synchronized, circular motion. The image reveals a minimalist world composed of pure black, pure white, and countless shades of gray in between. A vast, mirror-like expanse of gray water reflects the leaden sky, the two colors blending seamlessly, with almost no demarcation. In the center of the water, a simple marker, made from a few weathered bamboo poles, stands alone. A fisherman steers a narrow, dark-black wooden boat, its simple lines stained with a subtle matte sheen from years of use. With his back to the camera, the fisherman, hunched over, wears a wide, traditional bamboo hat woven from withered bamboo strips and a loose, dark-gray cotton-linen jacket. He gripped a long wooden oar tightly in both hands, each stroke appearing unhurried and rhythmic. The oar's impact on the water only stirred faint, rapidly expanding white ripples on the mirror-like surface. The boat moved slowly, at an almost meditative pace, in a perfect circle, circling the bamboo pole mark. The even, soft, diffuse light outlined the contours of everything without creating noticeable shadows, enhancing the flatness and abstraction of the painting. The overall atmosphere was imbued with the poetic and Zen-like qualities of an Eastern ink painting. An ultimate sense of tranquility, solitude, and eternity enveloped the entire scene.", + "seed": 26504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_15/index_5169_seed_26504_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": -0.029193812608718873, + "vlm_logprob": -0.029193812608718873, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.029193812608718873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.022787846624851227 + ], + "opposite_logprob_tokens": [ + -5.022788047790527 + ], + "logprob_reward": -0.022787846624851227, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.022787846624851227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1516876369714737, + 0.11511246860027313, + 0.042645253241062164, + 0.10613841563463211, + 0.09215720742940903, + 0.1036287322640419, + 0.10212834179401398, + 0.0443769209086895, + 0.06740962713956833, + 0.17471544444561005 + ], + "credit_map": [ + 0.0180303193628788, + 0.014998499304056168, + 0.026333576068282127, + 0.013545023277401924, + 0.015638276934623718, + 0.01261123176664114, + 0.013223156332969666, + 0.03730754926800728, + 0.01570463553071022, + 0.01322634145617485, + 0.015673592686653137, + 0.01069722045212984, + 0.010841209441423416, + 0.009435941465198994, + 0.008246938697993755, + 0.031286582350730896, + 0.005774214398115873, + 0.004841921851038933, + 0.0071600754745304585, + 0.005939285270869732, + 0.0046653542667627335, + 0.0029461104422807693, + 0.004316315520554781, + 0.007001975551247597, + 0.016818942502141, + 0.01185834314674139, + 0.02103229984641075, + 0.0075574638321995735, + 0.007461656350642443, + 0.014782636426389217, + 0.008258901536464691, + 0.01836819015443325, + 0.009784713387489319, + 0.01571073941886425, + 0.012790367938578129, + 0.011053881607949734, + 0.01381604839116335, + 0.008548427373170853, + 0.007639984600245953, + 0.012813041917979717, + 0.005685306619852781, + 0.030645424500107765, + 0.00564087787643075, + 0.005240294616669416, + 0.008089864626526833, + 0.03753087669610977, + 0.005036776885390282, + 0.005759308114647865, + 0.015365283004939556, + 0.020019004121422768, + 0.009617275558412075, + 0.008356300182640553, + 0.009245352819561958, + 0.017223812639713287, + 0.007189470808953047, + 0.015111844055354595, + 0.004398777149617672, + 0.006046016234904528, + 0.007362878415733576, + 0.0031755941454321146, + 0.006953860633075237, + 0.0062742591835558414, + 0.0026891082525253296, + 0.007476428523659706, + 0.007771007250994444, + 0.008472766727209091, + 0.017504893243312836, + 0.00413219491019845, + 0.008408638648688793, + 0.006997513119131327, + 0.0035302231553941965, + 0.010592388920485973, + 0.015114715322852135, + 0.02435285411775112, + 0.042917631566524506, + 0.01520113181322813, + 0.012021305039525032, + 0.0298048984259367, + 0.010799170471727848, + 0.024503741413354874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.022787846624851227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.020611152052879333 + ], + "opposite_logprob_tokens": [ + -3.895611047744751 + ], + "logprob_reward": -0.020611152052879333, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020611152052879333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2902814745903015, + 0.23212112486362457, + 0.05956729128956795, + 0.07182496786117554, + 0.10831983387470245, + 0.04492002725601196, + 0.0807376429438591, + 0.03218977153301239, + 0.04161280393600464, + 0.038425132632255554 + ], + "credit_map": [ + 0.04438737407326698, + 0.031924057751894, + 0.04696829244494438, + 0.02093111351132393, + 0.023855848237872124, + 0.03451795503497124, + 0.026708992198109627, + 0.06098778545856476, + 0.049093425273895264, + 0.03244742006063461, + 0.03239310905337334, + 0.012900964356958866, + 0.013850877061486244, + 0.028301598504185677, + 0.015769867226481438, + 0.047363847494125366, + 0.007547995075583458, + 0.006625404581427574, + 0.01013046782463789, + 0.006602591369301081, + 0.007695487234741449, + 0.005486331880092621, + 0.005970927886664867, + 0.009508081711828709, + 0.010638379491865635, + 0.010840627364814281, + 0.011855492368340492, + 0.006135738920420408, + 0.008989734575152397, + 0.005982242524623871, + 0.004314596299082041, + 0.01306814793497324, + 0.013461004942655563, + 0.019747929647564888, + 0.013334542512893677, + 0.013106911443173885, + 0.014092831872403622, + 0.009363858960568905, + 0.009445960633456707, + 0.01576678827404976, + 0.006127506960183382, + 0.00696761067956686, + 0.00690646143630147, + 0.004055139143019915, + 0.0037097556050866842, + 0.005807224661111832, + 0.003161380998790264, + 0.008184942416846752, + 0.013295486569404602, + 0.016965389251708984, + 0.006414658855646849, + 0.003986846189945936, + 0.013008413836359978, + 0.01685195602476597, + 0.003928111866116524, + 0.006286775693297386, + 0.0038403731305152178, + 0.004469335079193115, + 0.0048357294872403145, + 0.0031769275665283203, + 0.005519745405763388, + 0.0035601535346359015, + 0.0030324177350848913, + 0.0037550886627286673, + 0.0038242877926677465, + 0.004808309022337198, + 0.009210646152496338, + 0.003739769570529461, + 0.005350402556359768, + 0.004291946534067392, + 0.003976133186370134, + 0.006411305628716946, + 0.0036088789347559214, + 0.005288749933242798, + 0.005155377089977264, + 0.003428623778745532, + 0.0026025811675935984, + 0.008180425502359867, + 0.002780376235023141, + 0.007380117196589708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.020611152052879333 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6208692789077759, + 0.06666487455368042, + 0.0529414638876915, + 0.031130196526646614, + 0.08421488851308823, + 0.028922807425260544, + 0.02734304778277874, + 0.044018663465976715, + 0.024157175794243813, + 0.019737696275115013 + ], + "credit_map": [ + 0.11753485351800919, + 0.07290493696928024, + 0.1092841848731041, + 0.060694433748722076, + 0.057714954018592834, + 0.061646897345781326, + 0.044037532061338425, + 0.09705140441656113, + 0.012317215092480183, + 0.008585583418607712, + 0.009769350290298462, + 0.0047489674761891365, + 0.005777318496257067, + 0.008939772844314575, + 0.005139836575835943, + 0.011386819183826447, + 0.0074234320782125, + 0.006229789927601814, + 0.009390313178300858, + 0.006207450293004513, + 0.005884163547307253, + 0.00546489842236042, + 0.004358895123004913, + 0.007982512004673481, + 0.004921258892863989, + 0.0035107641015201807, + 0.00626117829233408, + 0.0026625425089150667, + 0.002109592081978917, + 0.004252404905855656, + 0.0028083797078579664, + 0.004604075103998184, + 0.00834899302572012, + 0.013122323900461197, + 0.01033951248973608, + 0.010880793444812298, + 0.015055716969072819, + 0.00788345281034708, + 0.007584563922137022, + 0.01099952682852745, + 0.003129401244223118, + 0.0043892813846468925, + 0.002775792730972171, + 0.0019752378575503826, + 0.005105132237076759, + 0.005932690110057592, + 0.001554103335365653, + 0.004061167594045401, + 0.0037845910992473364, + 0.0036586758214980364, + 0.00291179190389812, + 0.0023797652684152126, + 0.005686948075890541, + 0.0032553360797464848, + 0.002114869887009263, + 0.0035510663874447346, + 0.004069373477250338, + 0.00499103544279933, + 0.005014317110180855, + 0.004856433719396591, + 0.005729726515710354, + 0.005876174662262201, + 0.004965687170624733, + 0.008515913970768452, + 0.0018305175472050905, + 0.0037276900839060545, + 0.0036500857677310705, + 0.0021742384415119886, + 0.0024194326251745224, + 0.005062390584498644, + 0.0021347154397517443, + 0.003158104605972767, + 0.002136235823854804, + 0.002811718499287963, + 0.005114913918077946, + 0.001304415985941887, + 0.0015521495370194316, + 0.0025128622073680162, + 0.0008820825023576617, + 0.003423317801207304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0110772168263793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.012542181648314 + ], + "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": [ + 0.3187306225299835, + 0.12904997169971466, + 0.0657663568854332, + 0.04896912723779678, + 0.07917768508195877, + 0.04420486465096474, + 0.12573213875293732, + 0.048193130642175674, + 0.051305994391441345, + 0.08887002617120743 + ], + "credit_map": [ + 0.05752026289701462, + 0.042920518666505814, + 0.053921639919281006, + 0.018784979358315468, + 0.029647085815668106, + 0.0527859665453434, + 0.022945621982216835, + 0.040204551070928574, + 0.014226727187633514, + 0.01261143572628498, + 0.018342847004532814, + 0.010366245172917843, + 0.0134927062317729, + 0.026651453226804733, + 0.009647483006119728, + 0.02371107041835785, + 0.011383289471268654, + 0.008284424431622028, + 0.01300253439694643, + 0.005827673710882664, + 0.007928943261504173, + 0.004605712369084358, + 0.005421873647719622, + 0.009311908856034279, + 0.008492646738886833, + 0.0054749995470047, + 0.009211083874106407, + 0.003184880595654249, + 0.003242906415835023, + 0.007308525033295155, + 0.002989624859765172, + 0.00906446110457182, + 0.012041006237268448, + 0.01299198716878891, + 0.012301855720579624, + 0.007026724983006716, + 0.007018475793302059, + 0.010305500589311123, + 0.0075112986378371716, + 0.009980839677155018, + 0.004230862949043512, + 0.009015694260597229, + 0.0033996354322880507, + 0.0018107774667441845, + 0.0072305756621062756, + 0.012099510058760643, + 0.001976148923859, + 0.004441664554178715, + 0.014445565640926361, + 0.020575501024723053, + 0.01041792705655098, + 0.008533592335879803, + 0.0250928346067667, + 0.021499289199709892, + 0.010452925227582455, + 0.014714508317410946, + 0.004534388892352581, + 0.006024755537509918, + 0.008852537721395493, + 0.003450086573138833, + 0.0071738543920218945, + 0.005899007432162762, + 0.0038620298728346825, + 0.008396473713219166, + 0.004044945351779461, + 0.006145599763840437, + 0.008024253882467747, + 0.006687643937766552, + 0.005703471601009369, + 0.00610797293484211, + 0.006341639906167984, + 0.008250468410551548, + 0.008671571500599384, + 0.01459906343370676, + 0.01950964331626892, + 0.007272548042237759, + 0.006232242565602064, + 0.01523553766310215, + 0.003808858571574092, + 0.013540562242269516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.012542181648314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.0789506658911705 + ], + "opposite_logprob_tokens": [ + -2.5789506435394287 + ], + "logprob_reward": -0.0789506658911705, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789506658911705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34087827801704407, + 0.05651657655835152, + 0.08409460633993149, + 0.04584923014044762, + 0.0864047110080719, + 0.11781112849712372, + 0.12660132348537445, + 0.037689752876758575, + 0.05726969987154007, + 0.046884696930646896 + ], + "credit_map": [ + 0.0448031909763813, + 0.03128581494092941, + 0.05652597174048424, + 0.03962978348135948, + 0.03010174073278904, + 0.044236358255147934, + 0.030979406088590622, + 0.06331601738929749, + 0.008952880278229713, + 0.005209224298596382, + 0.00864128116518259, + 0.006015331018716097, + 0.004534031730145216, + 0.006990522146224976, + 0.00551624596118927, + 0.010657057166099548, + 0.010179871693253517, + 0.008460749872028828, + 0.01672276109457016, + 0.010157844983041286, + 0.009217621758580208, + 0.007881722413003445, + 0.00914523471146822, + 0.012328801676630974, + 0.005324366968125105, + 0.005816429853439331, + 0.007706753443926573, + 0.006514781154692173, + 0.003872697241604328, + 0.0045075505040585995, + 0.004818219691514969, + 0.007288435939699411, + 0.013839601539075375, + 0.012716875411570072, + 0.018264126032590866, + 0.006174597889184952, + 0.006793648935854435, + 0.011594112031161785, + 0.005127528216689825, + 0.011894220486283302, + 0.016410378739237785, + 0.01733100228011608, + 0.013972446322441101, + 0.010797158814966679, + 0.00885048508644104, + 0.014776777476072311, + 0.0091819753870368, + 0.02649090811610222, + 0.013212170451879501, + 0.016857381910085678, + 0.011977886781096458, + 0.010332946665585041, + 0.02290419302880764, + 0.017171265557408333, + 0.010401660576462746, + 0.023743826895952225, + 0.0039347149431705475, + 0.0043809241615235806, + 0.004687194712460041, + 0.0034178642090409994, + 0.0068532731384038925, + 0.004860401153564453, + 0.0032695054542273283, + 0.006285874638706446, + 0.005747666582465172, + 0.0068143876269459724, + 0.013410717248916626, + 0.004676530137658119, + 0.006375730037689209, + 0.005965122953057289, + 0.004501649644225836, + 0.009777896106243134, + 0.004907920025289059, + 0.006470454391092062, + 0.008290768601000309, + 0.005761387292295694, + 0.005992326885461807, + 0.005749971140176058, + 0.004459703341126442, + 0.005252164322882891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0789506658911705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot captured from a high-angle, slightly bird's-eye view from a drone, the camera slowly and steadily follows the subject in synchronized, circular motion. The image reveals a minimalist world composed of pure black, pure white, and countless shades of gray in between. A vast, mirror-like expanse of gray water reflects the leaden sky, the two colors blending seamlessly, with almost no demarcation. In the center of the water, a simple marker, made from a few weathered bamboo poles, stands alone. A fisherman steers a narrow, dark-black wooden boat, its simple lines stained with a subtle matte sheen from years of use. With his back to the camera, the fisherman, hunched over, wears a wide, traditional bamboo hat woven from withered bamboo strips and a loose, dark-gray cotton-linen jacket. He gripped a long wooden oar tightly in both hands, each stroke appearing unhurried and rhythmic. The oar's impact on the water only stirred faint, rapidly expanding white ripples on the mirror-like surface. The boat moved slowly, at an almost meditative pace, in a perfect circle, circling the bamboo pole mark. The even, soft, diffuse light outlined the contours of everything without creating noticeable shadows, enhancing the flatness and abstraction of the painting. The overall atmosphere was imbued with the poetic and Zen-like qualities of an Eastern ink painting. An ultimate sense of tranquility, solitude, and eternity enveloped the entire scene.", + "seed": 26505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_15/index_5169_seed_26505_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": -0.038792801462113856, + "vlm_logprob": -0.038792801462113856, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.038792801462113856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.03573806211352348 + ], + "opposite_logprob_tokens": [ + -3.785737991333008 + ], + "logprob_reward": -0.03573806211352348, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03573806211352348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15641021728515625, + 0.14484873414039612, + 0.07445447891950607, + 0.2022942751646042, + 0.21485254168510437, + 0.0873173251748085, + 0.017289482057094574, + 0.02666115202009678, + 0.04781213402748108, + 0.028059586882591248 + ], + "credit_map": [ + 0.021929770708084106, + 0.026891514658927917, + 0.021212173625826836, + 0.023571506142616272, + 0.010120999999344349, + 0.018440980464220047, + 0.011324085295200348, + 0.0229191891849041, + 0.012328138574957848, + 0.010283675976097584, + 0.045706626027822495, + 0.015857290476560593, + 0.006482287310063839, + 0.00756046362221241, + 0.011282685212790966, + 0.03534755855798721, + 0.009709148667752743, + 0.010511565022170544, + 0.009844339452683926, + 0.009630230255424976, + 0.00852889847010374, + 0.006908752955496311, + 0.0075555662624537945, + 0.0117659792304039, + 0.04612550884485245, + 0.051973555237054825, + 0.022762805223464966, + 0.013179698958992958, + 0.012193011119961739, + 0.027209699153900146, + 0.012990745715796947, + 0.015859242528676987, + 0.07331489026546478, + 0.08026670664548874, + 0.012321120128035545, + 0.009585099294781685, + 0.013823503628373146, + 0.008128516376018524, + 0.008743409998714924, + 0.008669300936162472, + 0.015416969545185566, + 0.014358675107359886, + 0.011759085580706596, + 0.007493563462048769, + 0.015497032552957535, + 0.008578459732234478, + 0.006783133838325739, + 0.007430416997522116, + 0.0032455846667289734, + 0.002315614139661193, + 0.002263532718643546, + 0.0016139588551595807, + 0.0019124101381748915, + 0.0028095741290599108, + 0.0010742502054199576, + 0.0020545574370771646, + 0.003929773811250925, + 0.005632144398987293, + 0.003061269875615835, + 0.0021965077612549067, + 0.003882028628140688, + 0.0038900687359273434, + 0.0016927969409152865, + 0.002376559190452099, + 0.005025774706155062, + 0.0073692686855793, + 0.006928877905011177, + 0.004060893785208464, + 0.008353537879884243, + 0.008006630465388298, + 0.002830497920513153, + 0.005236656405031681, + 0.0026535498909652233, + 0.003394713392481208, + 0.003058935981243849, + 0.005508590955287218, + 0.0032806252129375935, + 0.0026558823883533478, + 0.0030919916462153196, + 0.004415298346430063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03573806211352348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.05498363450169563 + ], + "opposite_logprob_tokens": [ + -2.929983615875244 + ], + "logprob_reward": -0.05498363450169563, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498363450169563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3431393802165985, + 0.1463656723499298, + 0.056965421885252, + 0.08989608287811279, + 0.11334526538848877, + 0.060228604823350906, + 0.014906813390552998, + 0.0602877177298069, + 0.06121120974421501, + 0.05365380644798279 + ], + "credit_map": [ + 0.06021224334836006, + 0.0629030168056488, + 0.05026214197278023, + 0.04849942773580551, + 0.02452048659324646, + 0.04057391732931137, + 0.023316971957683563, + 0.032851167023181915, + 0.01705767773091793, + 0.011737930588424206, + 0.03691810742020607, + 0.013201179914176464, + 0.008825503289699554, + 0.010499531403183937, + 0.012113978154957294, + 0.036011770367622375, + 0.0071453857235610485, + 0.007578159682452679, + 0.00838119350373745, + 0.009166273288428783, + 0.005243927706032991, + 0.00539338868111372, + 0.006302088964730501, + 0.007755008060485125, + 0.017285706475377083, + 0.02112545818090439, + 0.010118069127202034, + 0.0065770335495471954, + 0.006809357088059187, + 0.006577595137059689, + 0.01070973090827465, + 0.010693138465285301, + 0.027707776054739952, + 0.028668351471424103, + 0.015552139841020107, + 0.012372525408864021, + 0.011352146044373512, + 0.0043744659051299095, + 0.006629460025578737, + 0.006688394583761692, + 0.011498368345201015, + 0.010586961172521114, + 0.00848530512303114, + 0.0048014274798333645, + 0.012489148415625095, + 0.004666926804929972, + 0.003237329190596938, + 0.004463139455765486, + 0.002081480575725436, + 0.0021262632217258215, + 0.0015822036657482386, + 0.0014978029066696763, + 0.002355081494897604, + 0.0027329926379024982, + 0.0012139197206124663, + 0.0013170684687793255, + 0.006455312017351389, + 0.013871791772544384, + 0.005527939647436142, + 0.005128972232341766, + 0.013927880674600601, + 0.0063205440528690815, + 0.004355615004897118, + 0.004699665121734142, + 0.004439517855644226, + 0.007578446064144373, + 0.007521859370172024, + 0.007929150015115738, + 0.012995639815926552, + 0.010763630270957947, + 0.003902183845639229, + 0.006080785766243935, + 0.006839079316705465, + 0.007035300601273775, + 0.006179043557494879, + 0.007692506071180105, + 0.007462376728653908, + 0.006870198529213667, + 0.004415767267346382, + 0.007159536704421043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05498363450169563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.05495012179017067 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.05495012179017067, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495012179017067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08241212368011475, + 0.049918439239263535, + 0.058737751096487045, + 0.5286139845848083, + 0.09417930245399475, + 0.032006315886974335, + 0.021569309756159782, + 0.03490760549902916, + 0.049733661115169525, + 0.04792153835296631 + ], + "credit_map": [ + 0.009295063093304634, + 0.009641917422413826, + 0.012465450912714005, + 0.018763521686196327, + 0.005264397244900465, + 0.010065806098282337, + 0.0068225194700062275, + 0.010093437507748604, + 0.0051203565672039986, + 0.0039780777879059315, + 0.012499812059104443, + 0.0074798790737986565, + 0.0032911256421357393, + 0.003754178062081337, + 0.004198937676846981, + 0.009596070274710655, + 0.00942098256200552, + 0.006602680776268244, + 0.007692164741456509, + 0.006296319887042046, + 0.004881567787379026, + 0.00868843775242567, + 0.006353921722620726, + 0.008801673538982868, + 0.10027530044317245, + 0.11706992238759995, + 0.06990066915750504, + 0.03211330994963646, + 0.03152163699269295, + 0.08654375374317169, + 0.032185327261686325, + 0.059004046022892, + 0.02536885067820549, + 0.02812114916741848, + 0.009335681796073914, + 0.007923461496829987, + 0.006712235510349274, + 0.003365181852132082, + 0.0065416013821959496, + 0.006811139173805714, + 0.003917183727025986, + 0.005331265274435282, + 0.008173475041985512, + 0.0027605195064097643, + 0.005985973868519068, + 0.0021194959990680218, + 0.0015440331771969795, + 0.00217436533421278, + 0.0046495841816067696, + 0.0030090545769780874, + 0.00353133212774992, + 0.0013916189782321453, + 0.002468858379870653, + 0.002401170553639531, + 0.001237658434547484, + 0.0028800310101360083, + 0.003141195746138692, + 0.007752607110887766, + 0.0036754263564944267, + 0.002767829457297921, + 0.006475390400737524, + 0.00532161770388484, + 0.002539680805057287, + 0.0032338560558855534, + 0.005666657350957394, + 0.007359267212450504, + 0.008483227342367172, + 0.00387296499684453, + 0.008347313851118088, + 0.009791548363864422, + 0.0024594853166490793, + 0.00375319248996675, + 0.0054685212671756744, + 0.007258969824761152, + 0.005753050558269024, + 0.003882631193846464, + 0.007859377190470695, + 0.009274041280150414, + 0.0033105863258242607, + 0.005114358384162188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.05495012179017067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.005253794603049755 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253794603049755, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253794603049755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07905871421098709, + 0.0796673521399498, + 0.02705340087413788, + 0.10187831521034241, + 0.08061040192842484, + 0.021426504477858543, + 0.021199151873588562, + 0.05581802502274513, + 0.08582966029644012, + 0.4474584460258484 + ], + "credit_map": [ + 0.010947797447443008, + 0.009507924318313599, + 0.011212417855858803, + 0.0132066048681736, + 0.007828119210898876, + 0.007860478945076466, + 0.008678263053297997, + 0.009817108511924744, + 0.006438888609409332, + 0.004446986131370068, + 0.022435404360294342, + 0.009735521860420704, + 0.00665662158280611, + 0.004180125892162323, + 0.007640316616743803, + 0.018133495002985, + 0.004211389925330877, + 0.00292029301635921, + 0.0035328161902725697, + 0.0032812957651913166, + 0.0022146813571453094, + 0.003968468401581049, + 0.0030723027884960175, + 0.0038521536625921726, + 0.01879604533314705, + 0.021779397502541542, + 0.013071293942630291, + 0.006952423602342606, + 0.006291344296187162, + 0.017385678365826607, + 0.00673397071659565, + 0.01086816843599081, + 0.02258262224495411, + 0.02346760779619217, + 0.008850112557411194, + 0.00614306703209877, + 0.006067830137908459, + 0.00411636708304286, + 0.00439494289457798, + 0.004987852647900581, + 0.003232586430385709, + 0.0031393037643283606, + 0.0035991778131574392, + 0.002188315149396658, + 0.00415332056581974, + 0.0019119917415082455, + 0.0013716884423047304, + 0.0018301201052963734, + 0.003369756508618593, + 0.002847355557605624, + 0.0023142157588154078, + 0.0013087308034300804, + 0.0033476343378424644, + 0.004820163361728191, + 0.0013151346938684583, + 0.0018761609680950642, + 0.006571526639163494, + 0.012705797329545021, + 0.004955227952450514, + 0.006563796661794186, + 0.00859820656478405, + 0.0063339713960886, + 0.004545111674815416, + 0.005544389132410288, + 0.008428257890045643, + 0.01578010991215706, + 0.010901860892772675, + 0.008544960990548134, + 0.016440844163298607, + 0.013391556218266487, + 0.004912105854600668, + 0.007429968100041151, + 0.04669322445988655, + 0.08168010413646698, + 0.061954062432050705, + 0.047731149941682816, + 0.07130002975463867, + 0.07306947559118271, + 0.03237532451748848, + 0.032655052840709686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.005253794603049755 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.043038394302129745 + ], + "opposite_logprob_tokens": [ + -3.1680383682250977 + ], + "logprob_reward": -0.043038394302129745, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043038394302129745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12569959461688995, + 0.13976940512657166, + 0.09986001253128052, + 0.0976627916097641, + 0.1520232856273651, + 0.13572092354297638, + 0.029822800308465958, + 0.04484837129712105, + 0.11958933621644974, + 0.05500354245305061 + ], + "credit_map": [ + 0.017714451998472214, + 0.010848079808056355, + 0.019048213958740234, + 0.020790619775652885, + 0.009253144264221191, + 0.017929673194885254, + 0.013033909723162651, + 0.01708149164915085, + 0.014390869997441769, + 0.014179086312651634, + 0.025927308946847916, + 0.022792506963014603, + 0.009032008238136768, + 0.012754758819937706, + 0.012596496380865574, + 0.02809636853635311, + 0.011602240614593029, + 0.01266186498105526, + 0.012842213734984398, + 0.012181338854134083, + 0.013405601494014263, + 0.008832238614559174, + 0.01160342711955309, + 0.016731083393096924, + 0.011630863882601261, + 0.011869042180478573, + 0.017242861911654472, + 0.007708503864705563, + 0.007644070778042078, + 0.01815197803080082, + 0.00794326700270176, + 0.015472201630473137, + 0.04849233478307724, + 0.05498300492763519, + 0.008713273331522942, + 0.009261122904717922, + 0.010589280165731907, + 0.005453675519675016, + 0.008607384748756886, + 0.005923214368522167, + 0.01669660583138466, + 0.01886182650923729, + 0.036387037485837936, + 0.013322248123586178, + 0.019702568650245667, + 0.00903324969112873, + 0.0088046221062541, + 0.012912754900753498, + 0.004232997074723244, + 0.006072354037314653, + 0.004035673104226589, + 0.002508319914340973, + 0.005186410155147314, + 0.0034151002764701843, + 0.0018368820892646909, + 0.0025350647047162056, + 0.007412105333060026, + 0.006515834480524063, + 0.004899114370346069, + 0.0021354197524487972, + 0.006792651955038309, + 0.010349776595830917, + 0.0019611765164881945, + 0.004782292991876602, + 0.013171945698559284, + 0.01957850717008114, + 0.024723807349801064, + 0.00962994433939457, + 0.015427562408149242, + 0.01741126924753189, + 0.008298390544950962, + 0.011347907595336437, + 0.004977515432983637, + 0.01034107431769371, + 0.008656322956085205, + 0.006681426893919706, + 0.006741277873516083, + 0.008598856627941132, + 0.0034268293529748917, + 0.005580239929258823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.043038394302129745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot captured from a high-angle, slightly bird's-eye view from a drone, the camera slowly and steadily follows the subject in synchronized, circular motion. The image reveals a minimalist world composed of pure black, pure white, and countless shades of gray in between. A vast, mirror-like expanse of gray water reflects the leaden sky, the two colors blending seamlessly, with almost no demarcation. In the center of the water, a simple marker, made from a few weathered bamboo poles, stands alone. A fisherman steers a narrow, dark-black wooden boat, its simple lines stained with a subtle matte sheen from years of use. With his back to the camera, the fisherman, hunched over, wears a wide, traditional bamboo hat woven from withered bamboo strips and a loose, dark-gray cotton-linen jacket. He gripped a long wooden oar tightly in both hands, each stroke appearing unhurried and rhythmic. The oar's impact on the water only stirred faint, rapidly expanding white ripples on the mirror-like surface. The boat moved slowly, at an almost meditative pace, in a perfect circle, circling the bamboo pole mark. The even, soft, diffuse light outlined the contours of everything without creating noticeable shadows, enhancing the flatness and abstraction of the painting. The overall atmosphere was imbued with the poetic and Zen-like qualities of an Eastern ink painting. An ultimate sense of tranquility, solitude, and eternity enveloped the entire scene.", + "seed": 26506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_15/index_5169_seed_26506_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": -0.0730981208384037, + "vlm_logprob": -0.0730981208384037, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.0730981208384037, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.028577767312526703 + ], + "opposite_logprob_tokens": [ + -4.15357780456543 + ], + "logprob_reward": -0.028577767312526703, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.028577767312526703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08735445886850357, + 0.04314909130334854, + 0.030014341697096825, + 0.10031480342149734, + 0.04209381714463234, + 0.022587528452277184, + 0.09015300124883652, + 0.0318286158144474, + 0.09364093095064163, + 0.45886343717575073 + ], + "credit_map": [ + 0.008587408810853958, + 0.013334743678569794, + 0.017656706273555756, + 0.007736682891845703, + 0.00972975417971611, + 0.00969778373837471, + 0.009203335270285606, + 0.011408036574721336, + 0.003960164729505777, + 0.0061609563417732716, + 0.010049230419099331, + 0.0032884664833545685, + 0.004124533850699663, + 0.005846282932907343, + 0.00458208192139864, + 0.005137372761964798, + 0.0036861051339656115, + 0.005463609471917152, + 0.004821191541850567, + 0.0021055929828435183, + 0.0026679816655814648, + 0.0044154510833323, + 0.0028456347063183784, + 0.00400877557694912, + 0.007706308271735907, + 0.013495084829628468, + 0.015602302737534046, + 0.00625008437782526, + 0.013207010924816132, + 0.01887892372906208, + 0.00788034312427044, + 0.0172947458922863, + 0.004708751104772091, + 0.007509893737733364, + 0.008369043469429016, + 0.002859829692170024, + 0.004269974306225777, + 0.005919496528804302, + 0.0032067180145531893, + 0.005250107496976852, + 0.002438526600599289, + 0.004833024926483631, + 0.0022466762457042933, + 0.0018056245753541589, + 0.0025866744108498096, + 0.00409059040248394, + 0.002019440056756139, + 0.0025669720489531755, + 0.007053133100271225, + 0.016813015565276146, + 0.011094029061496258, + 0.007765534333884716, + 0.013356379233300686, + 0.016646867617964745, + 0.008529802784323692, + 0.008894244208931923, + 0.003195197321474552, + 0.0073728482238948345, + 0.004032979253679514, + 0.002460142830386758, + 0.0034650592133402824, + 0.006014110986143351, + 0.002017863793298602, + 0.003270411863923073, + 0.009199658408761024, + 0.012377822771668434, + 0.015996510162949562, + 0.01150800846517086, + 0.010788793675601482, + 0.011017724871635437, + 0.010293365456163883, + 0.012459047138690948, + 0.048283908516168594, + 0.10856590420007706, + 0.14109276235103607, + 0.08468394726514816, + 0.02011122554540634, + 0.02512829378247261, + 0.004378368146717548, + 0.026619046926498413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.028577767312526703 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.25213417410850525 + ], + "opposite_logprob_tokens": [ + -1.5021342039108276 + ], + "logprob_reward": -0.25213417410850525, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213417410850525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19196245074272156, + 0.09724267572164536, + 0.0503060445189476, + 0.035777267068624496, + 0.03789588436484337, + 0.04391526058316231, + 0.11891564726829529, + 0.07721369713544846, + 0.05409861356019974, + 0.2926725149154663 + ], + "credit_map": [ + 0.016430163756012917, + 0.029853997752070427, + 0.04087577015161514, + 0.019311293959617615, + 0.018976837396621704, + 0.020193247124552727, + 0.023459166288375854, + 0.022861965000629425, + 0.011336083523929119, + 0.014440882951021194, + 0.018013732507824898, + 0.008351802825927734, + 0.007941408082842827, + 0.012864074669778347, + 0.01063478086143732, + 0.013659913092851639, + 0.0056054978631436825, + 0.006647000554949045, + 0.008136293850839138, + 0.004240593872964382, + 0.005282023921608925, + 0.0066820308566093445, + 0.006419154349714518, + 0.007293445989489555, + 0.0050428383983671665, + 0.004522419068962336, + 0.005314038135111332, + 0.002338957507163286, + 0.00351733504794538, + 0.006123617757111788, + 0.0044178529642522335, + 0.0045002056285738945, + 0.0052110287360847, + 0.004692554473876953, + 0.006220455281436443, + 0.005370353814214468, + 0.0044444045051932335, + 0.003031968604773283, + 0.003541716141626239, + 0.005383402109146118, + 0.005673385690897703, + 0.005507197696715593, + 0.0064742304384708405, + 0.004016593564301729, + 0.003930799197405577, + 0.009113126434385777, + 0.003716481616720557, + 0.005483448505401611, + 0.008631768636405468, + 0.03435756266117096, + 0.008550274185836315, + 0.006264758296310902, + 0.013181169517338276, + 0.031455013900995255, + 0.008541916497051716, + 0.007933182641863823, + 0.005724439863115549, + 0.020103609189391136, + 0.0066319420002400875, + 0.005158453714102507, + 0.008471273817121983, + 0.01937600038945675, + 0.004843220114707947, + 0.006904755253344774, + 0.004585105460137129, + 0.005785232409834862, + 0.008193114772439003, + 0.00565593084320426, + 0.011006904765963554, + 0.005862356163561344, + 0.006683437153697014, + 0.006326533854007721, + 0.01664041541516781, + 0.02504599653184414, + 0.11660070717334747, + 0.07907366007566452, + 0.009752841666340828, + 0.011609739623963833, + 0.008370045572519302, + 0.025579124689102173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.25213417410850525 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.02635536715388298 + ], + "opposite_logprob_tokens": [ + -3.651355266571045 + ], + "logprob_reward": -0.02635536715388298, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635536715388298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24321845173835754, + 0.2680518925189972, + 0.04813208058476448, + 0.07462332397699356, + 0.03256354108452797, + 0.017699483782052994, + 0.059144385159015656, + 0.08358414471149445, + 0.0412374846637249, + 0.13174520432949066 + ], + "credit_map": [ + 0.02634298801422119, + 0.038627635687589645, + 0.05086018145084381, + 0.025398876518011093, + 0.018912049010396004, + 0.02617977187037468, + 0.02846471033990383, + 0.028432225808501244, + 0.028119957074522972, + 0.0364995039999485, + 0.04250943660736084, + 0.01849776692688465, + 0.02350321225821972, + 0.05619826167821884, + 0.021431542932987213, + 0.0412922278046608, + 0.004770577419549227, + 0.00593961076810956, + 0.009189341217279434, + 0.0035953903570771217, + 0.00436391681432724, + 0.009876856580376625, + 0.0037752881180495024, + 0.006621099077165127, + 0.007161563728004694, + 0.011202685534954071, + 0.014315991662442684, + 0.004682581406086683, + 0.007140459958463907, + 0.012753250077366829, + 0.005205586086958647, + 0.012161200866103172, + 0.0026792159769684076, + 0.0044074696488678455, + 0.004277634434401989, + 0.0032270653173327446, + 0.004648858681321144, + 0.006069387309253216, + 0.0030195889994502068, + 0.004234321415424347, + 0.001548235071823001, + 0.0023383551742881536, + 0.0018784204730764031, + 0.0022965846583247185, + 0.0025918264873325825, + 0.002794358180835843, + 0.0022562406957149506, + 0.0019954631570726633, + 0.004341740161180496, + 0.011568651534616947, + 0.0067528835497796535, + 0.00553752388805151, + 0.008150609210133553, + 0.010370015166699886, + 0.006377944722771645, + 0.0060450173914432526, + 0.003944887313991785, + 0.033463142812252045, + 0.0023115279618650675, + 0.0026026524137705564, + 0.004033244680613279, + 0.03356247395277023, + 0.0016872399719431996, + 0.001978978281840682, + 0.003744681365787983, + 0.006896089296787977, + 0.005082874558866024, + 0.004286859184503555, + 0.004074505530297756, + 0.005309146363288164, + 0.00570627162232995, + 0.006137056276202202, + 0.029998507350683212, + 0.05108669400215149, + 0.013032829388976097, + 0.0054990556091070175, + 0.008503149263560772, + 0.013862581923604012, + 0.0038216107059270144, + 0.00594077305868268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.02635536715388298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.00978577509522438 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.00978577509522438, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978577509522438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09960208833217621, + 0.09404335170984268, + 0.03118479624390602, + 0.09527408331632614, + 0.07310840487480164, + 0.023145342245697975, + 0.09685149043798447, + 0.0891718715429306, + 0.05260320380330086, + 0.3450153172016144 + ], + "credit_map": [ + 0.012308560311794281, + 0.016250329092144966, + 0.016283098608255386, + 0.009857685305178165, + 0.008780443109571934, + 0.013659232296049595, + 0.009822724387049675, + 0.012640026398003101, + 0.008538401685655117, + 0.011815357953310013, + 0.01708642952144146, + 0.00853808969259262, + 0.007084075827151537, + 0.01579548977315426, + 0.01299548614770174, + 0.01219002902507782, + 0.003680186113342643, + 0.004397072363644838, + 0.004509939346462488, + 0.002676546573638916, + 0.0026454173494130373, + 0.003936804831027985, + 0.0038222046568989754, + 0.005516627803444862, + 0.007844077423214912, + 0.012339274398982525, + 0.016045182943344116, + 0.006409584544599056, + 0.011895242147147655, + 0.016575656831264496, + 0.00714190024882555, + 0.01702316664159298, + 0.007946056313812733, + 0.012246709316968918, + 0.013986376114189625, + 0.0064560892060399055, + 0.007774794939905405, + 0.0078030419535934925, + 0.007481861859560013, + 0.009413478896021843, + 0.0024634641595184803, + 0.0029566350858658552, + 0.004069128073751926, + 0.002330502960830927, + 0.0026853757444769144, + 0.003078416222706437, + 0.002685412298887968, + 0.0028764125891029835, + 0.004913047421723604, + 0.031011095270514488, + 0.006203576922416687, + 0.004497590009123087, + 0.008134258911013603, + 0.030911404639482498, + 0.005162863060832024, + 0.006017656531184912, + 0.00796718243509531, + 0.017823314294219017, + 0.009718573652207851, + 0.006370222195982933, + 0.011826983653008938, + 0.01595103181898594, + 0.007319296710193157, + 0.01219527143985033, + 0.006310324650257826, + 0.007729371078312397, + 0.006167454645037651, + 0.004836441483348608, + 0.009430396370589733, + 0.00490932259708643, + 0.006912339013069868, + 0.006307558622211218, + 0.03220072388648987, + 0.0714958980679512, + 0.10877154022455215, + 0.06793492287397385, + 0.01627890206873417, + 0.015567934140563011, + 0.0080950278788805, + 0.024670405313372612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.00978577509522438 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene viewed from a high-angle, slightly bird's-eye perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment composed primarily of black, white, and gray tones with a mirror-like gray water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a simple marker made of weathered bamboo poles visible in the center of the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fisherman wearing a traditional bamboo hat and a dark-gray jacket steering a narrow dark-black wooden boat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera follow the boat in a slow, steady, synchronized circular 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": [ + -0.048637520521879196 + ], + "opposite_logprob_tokens": [ + -3.048637628555298 + ], + "logprob_reward": -0.048637520521879196, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048637520521879196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034082867205142975, + 0.03199320659041405, + 0.027331681922078133, + 0.02974834479391575, + 0.022899528965353966, + 0.018112096935510635, + 0.031127341091632843, + 0.0606531985104084, + 0.06180022284388542, + 0.682251513004303 + ], + "credit_map": [ + 0.00469835102558136, + 0.005057924892753363, + 0.0055763195268809795, + 0.004207154270261526, + 0.003504251129925251, + 0.003921333700418472, + 0.0031982241198420525, + 0.003919308539479971, + 0.004266084171831608, + 0.006184423808008432, + 0.003373166546225548, + 0.0022066892124712467, + 0.0035737354774028063, + 0.005340151023119688, + 0.0031326794996857643, + 0.003916278015822172, + 0.002386645181104541, + 0.003694080049172044, + 0.004468945320695639, + 0.002786260563880205, + 0.002178031951189041, + 0.0039053522050380707, + 0.0031517588067799807, + 0.004760607611387968, + 0.0026744629722088575, + 0.0037726261653006077, + 0.005165841896086931, + 0.0016881704796105623, + 0.004113133065402508, + 0.005475222133100033, + 0.0020418406929820776, + 0.004817045293748379, + 0.00259751221165061, + 0.004781779367476702, + 0.004417757503688335, + 0.0018333707703277469, + 0.002178406808525324, + 0.0026466466952115297, + 0.0019214648054912686, + 0.002522587776184082, + 0.0016206681029871106, + 0.0020072453189641237, + 0.0030108557548373938, + 0.0020261669997125864, + 0.00245599914342165, + 0.002880336018279195, + 0.0017916855867952108, + 0.002319138729944825, + 0.0027548293583095074, + 0.005283771548420191, + 0.003409517230466008, + 0.0027989225927740335, + 0.005004245322197676, + 0.0057957288809120655, + 0.0031082499772310257, + 0.0029720747843384743, + 0.004086202941834927, + 0.019961155951023102, + 0.003235412761569023, + 0.0029327122028917074, + 0.005361376795917749, + 0.01854771003127098, + 0.002673432696610689, + 0.003855191869661212, + 0.006296692416071892, + 0.007724067196249962, + 0.006619278807193041, + 0.0051671890541911125, + 0.01696290448307991, + 0.005656642839312553, + 0.0068396320566535, + 0.006533816922456026, + 0.13272935152053833, + 0.24489472806453705, + 0.11035472899675369, + 0.04800090566277504, + 0.03851773217320442, + 0.06648522615432739, + 0.012135809287428856, + 0.02913302183151245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.048637520521879196 + } + ], + "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/0000131/group_2/index_tensor([6431, 6431])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_tensor([6431, 6431])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..db93bd7016d49c0dec2804c8adb815ef7490ce2e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_tensor([6431, 6431])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, rendered in the stark, beautifully minimalist style of Chinese ink painting, begins with a slow, deliberate downward tilt through a dense jungle of short, wide bamboo. The entire scene is a masterful performance of monochrome tones against a textured background of grayish-white rice paper, with soft, ethereal snowflakes silently drifting down. As the perspective descends, it reveals a subtle contrast between the pure, untouched white of the snow and the dark, damp surface of the bamboo. The broad leaves are depicted with bold, fluid strokes of ink, a deep, saturated black, their surfaces shimmering where the snow has yet to fall. The segmented stalks, rendered in varying shades of gray, their strong vertical forms contrast with the soft, horizontal layers of snow. The slow camera movement induces a contemplative focus, drawing the eye to the subtle interactions of snow and vegetation—a leaf gently bending under the weight of fresh powder, the sharp line where white snow meets black stem—creating an atmosphere of profound tranquility and minimalist beauty.", + "seed": 4149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_6431_seed_4149_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": -0.4190982226282358, + "vlm_logprob": -0.4190982226282358, + "vlm_yes_no_margin": 1.525, + "vlm_token_credit": -0.4190982226282358, + "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 video shot as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7376899123191833 + ], + "opposite_logprob_tokens": [ + -0.7376899123191833 + ], + "logprob_reward": -0.7376899123191833, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7376899123191833, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0685373842716217, + 0.03033345378935337, + 0.09523548930883408, + 0.03628629073500633, + 0.04757480323314667, + 0.06533393263816833, + 0.019763877615332603, + 0.017037395387887955, + 0.25685951113700867, + 0.3630378246307373 + ], + "credit_map": [ + 0.010592442005872726, + 0.010669400915503502, + 0.011333136819303036, + 0.007494079414755106, + 0.010798456147313118, + 0.006133877672255039, + 0.0063826581463217735, + 0.005133343394845724, + 0.004813159350305796, + 0.0044801984913647175, + 0.0037981723435223103, + 0.0030624682549387217, + 0.003675957443192601, + 0.004015759564936161, + 0.003804824547842145, + 0.0026829165872186422, + 0.018331294879317284, + 0.01000060886144638, + 0.009941055439412594, + 0.011542516760528088, + 0.015105659142136574, + 0.010343966074287891, + 0.00931033305823803, + 0.010660062544047832, + 0.007267852313816547, + 0.004211386200040579, + 0.005137778352946043, + 0.003978145774453878, + 0.00573298754170537, + 0.003586437553167343, + 0.0038986930157989264, + 0.0024730174336582422, + 0.0037805032916367054, + 0.0037117351312190294, + 0.0024457043036818504, + 0.002695506438612938, + 0.002974728588014841, + 0.012186056934297085, + 0.01617131382226944, + 0.0036092607770115137, + 0.01102488674223423, + 0.011819642968475819, + 0.005695992149412632, + 0.006142152938991785, + 0.008160519413650036, + 0.008441071026027203, + 0.00845907349139452, + 0.0055906083434820175, + 0.003253810340538621, + 0.001858755829744041, + 0.001811476657167077, + 0.0015548201045021415, + 0.002763202413916588, + 0.0028995461761951447, + 0.002783907577395439, + 0.002838362008333206, + 0.0025941766798496246, + 0.0014764959923923016, + 0.00235808826982975, + 0.0019590428564697504, + 0.002547111129388213, + 0.0014603693271055818, + 0.002135711954906583, + 0.0025064023211598396, + 0.01699005998671055, + 0.026603398844599724, + 0.04952777549624443, + 0.07233750075101852, + 0.02094740979373455, + 0.010337159037590027, + 0.029709136113524437, + 0.030407113954424858, + 0.039542652666568756, + 0.06597238034009933, + 0.08433804661035538, + 0.08808530867099762, + 0.029233606532216072, + 0.015005695633590221, + 0.024607712402939796, + 0.01625247485935688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.7376899123191833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0538731813430786 + ], + "opposite_logprob_tokens": [ + -0.4288731813430786 + ], + "logprob_reward": -1.0538731813430786, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538731813430786, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2638433873653412, + 0.06405801326036453, + 0.042544204741716385, + 0.04331904649734497, + 0.164926216006279, + 0.06331199407577515, + 0.0430728942155838, + 0.02649461105465889, + 0.16456268727779388, + 0.12386690080165863 + ], + "credit_map": [ + 0.0727374330163002, + 0.043487757444381714, + 0.026044929400086403, + 0.025139471516013145, + 0.028575139120221138, + 0.017532501369714737, + 0.03505491465330124, + 0.015271272510290146, + 0.009580975398421288, + 0.009433872066438198, + 0.007021639496088028, + 0.00650358060374856, + 0.0085698701441288, + 0.010232106782495975, + 0.004937684163451195, + 0.007778293918818235, + 0.008327377028763294, + 0.005104509182274342, + 0.004837012384086847, + 0.0031181820668280125, + 0.0071229045279324055, + 0.005315735004842281, + 0.0033690768759697676, + 0.005349410697817802, + 0.007934939116239548, + 0.005817284341901541, + 0.006147853564471006, + 0.00439255777746439, + 0.004676552955061197, + 0.004256946034729481, + 0.006823649164289236, + 0.0032692684326320887, + 0.010050098411738873, + 0.011039650067687035, + 0.006684625521302223, + 0.0056635248474776745, + 0.008399737067520618, + 0.051741670817136765, + 0.063580222427845, + 0.00776669243350625, + 0.007354677189141512, + 0.009702403098344803, + 0.007262480445206165, + 0.00425025774165988, + 0.005988882854580879, + 0.008312229067087173, + 0.013726750388741493, + 0.006714326329529285, + 0.007631581276655197, + 0.005008783657103777, + 0.004012431018054485, + 0.0032949340529739857, + 0.004717734642326832, + 0.006528204772621393, + 0.004304060712456703, + 0.007575168740004301, + 0.00636399956420064, + 0.0021143255289644003, + 0.00351798627525568, + 0.0021884702146053314, + 0.0019134539179503918, + 0.0025358612183481455, + 0.002746308222413063, + 0.005114210303872824, + 0.010769092477858067, + 0.01702655665576458, + 0.03013387694954872, + 0.04308567941188812, + 0.012413468211889267, + 0.0073024011217057705, + 0.021930109709501266, + 0.02190152183175087, + 0.013945497572422028, + 0.019670719280838966, + 0.02443452551960945, + 0.02672511152923107, + 0.011857411824166775, + 0.006234632804989815, + 0.009477992542088032, + 0.011521019972860813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -1.0538731813430786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097731068730354 + ], + "opposite_logprob_tokens": [ + -4.5110979080200195 + ], + "logprob_reward": -0.011097731068730354, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097731068730354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20670071244239807, + 0.09954280406236649, + 0.219980850815773, + 0.036442749202251434, + 0.06620007753372192, + 0.04133566841483116, + 0.03219671547412872, + 0.007962742820382118, + 0.1536087840795517, + 0.13602888584136963 + ], + "credit_map": [ + 0.06255725771188736, + 0.03521811217069626, + 0.02501729317009449, + 0.014844061806797981, + 0.023116840049624443, + 0.01714354380965233, + 0.01643439754843712, + 0.012369215488433838, + 0.022382883355021477, + 0.01619480922818184, + 0.010193252936005592, + 0.01129880826920271, + 0.00972199346870184, + 0.013208260759711266, + 0.007834894582629204, + 0.008707914501428604, + 0.038086965680122375, + 0.03216381371021271, + 0.018483078107237816, + 0.013020409271121025, + 0.056453194469213486, + 0.02477717213332653, + 0.018100667744874954, + 0.018895559012889862, + 0.005352023523300886, + 0.00445944257080555, + 0.005207162816077471, + 0.003486728062853217, + 0.005823287647217512, + 0.004251406993716955, + 0.005046652629971504, + 0.002816047752276063, + 0.0038342750631272793, + 0.004615723621100187, + 0.0036847221199423075, + 0.002677665324881673, + 0.00398494815453887, + 0.019282059744000435, + 0.02454371005296707, + 0.003576977178454399, + 0.0052016680128872395, + 0.003440934233367443, + 0.00397385424003005, + 0.002976116957142949, + 0.004531461745500565, + 0.007324567995965481, + 0.009829792194068432, + 0.004057273734360933, + 0.007777541875839233, + 0.0036097948905080557, + 0.0029356128070503473, + 0.003108238335698843, + 0.0037364130839705467, + 0.004499298520386219, + 0.002250454854220152, + 0.004279366694390774, + 0.0011966853635385633, + 0.0006446478655561805, + 0.001085870317183435, + 0.0007324860198423266, + 0.0009778771782293916, + 0.0007706666365265846, + 0.0007541296654380858, + 0.001800380414351821, + 0.011980689130723476, + 0.01659250818192959, + 0.03018120303750038, + 0.043308015912771225, + 0.01360613014549017, + 0.0061698914505541325, + 0.015950625762343407, + 0.015819739550352097, + 0.014828842133283615, + 0.0233116764575243, + 0.028302397578954697, + 0.031314313411712646, + 0.014106732793152332, + 0.005551009438931942, + 0.0090402252972126, + 0.009573696181178093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.011097731068730354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296184360980988 + ], + "opposite_logprob_tokens": [ + -2.237961769104004 + ], + "logprob_reward": -0.11296184360980988, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296184360980988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22815409302711487, + 0.04160661622881889, + 0.06932956725358963, + 0.04435563087463379, + 0.11174436658620834, + 0.1314719319343567, + 0.0483853854238987, + 0.018831389024853706, + 0.26938724517822266, + 0.036733850836753845 + ], + "credit_map": [ + 0.07227355986833572, + 0.03806072846055031, + 0.021816622465848923, + 0.01909639500081539, + 0.025928005576133728, + 0.020319275557994843, + 0.01908724009990692, + 0.011572243645787239, + 0.00821378082036972, + 0.006545810028910637, + 0.004971852991729975, + 0.0036274201702326536, + 0.004789539612829685, + 0.005216032732278109, + 0.003525812877342105, + 0.00471636513248086, + 0.011224756948649883, + 0.00969564076513052, + 0.005615105386823416, + 0.004911889787763357, + 0.015729695558547974, + 0.009069346822798252, + 0.006447003223001957, + 0.006636125035583973, + 0.007202479988336563, + 0.00627989461645484, + 0.006454484537243843, + 0.0040247682482004166, + 0.006691658403724432, + 0.004322847351431847, + 0.00605500303208828, + 0.0033244919031858444, + 0.008190309628844261, + 0.008634891360998154, + 0.005518097430467606, + 0.006594105623662472, + 0.006390267983078957, + 0.029950706288218498, + 0.039969466626644135, + 0.00649651512503624, + 0.021443981677293777, + 0.014785132370889187, + 0.013740522786974907, + 0.007830715738236904, + 0.009692533873021603, + 0.023920420557260513, + 0.033213287591934204, + 0.006845316383987665, + 0.00796103198081255, + 0.006035052705556154, + 0.0045808423310518265, + 0.004647726658731699, + 0.0070503526367247105, + 0.006480230484157801, + 0.005618621129542589, + 0.006011523772031069, + 0.003395999316126108, + 0.0017716075526550412, + 0.0025034097488969564, + 0.0018382353009656072, + 0.001817871816456318, + 0.002096747513860464, + 0.0021989501547068357, + 0.0032085655257105827, + 0.021279025822877884, + 0.026954611763358116, + 0.04883316531777382, + 0.07118505239486694, + 0.021284841001033783, + 0.011834417469799519, + 0.03285320848226547, + 0.03516288846731186, + 0.004022520966827869, + 0.0053420658223330975, + 0.006677860394120216, + 0.006504856515675783, + 0.0059656305238604546, + 0.0019075287273153663, + 0.0023635770194232464, + 0.003949807491153479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.11296184360980988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798684448003769 + ], + "opposite_logprob_tokens": [ + -1.804868459701538 + ], + "logprob_reward": -0.1798684448003769, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798684448003769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09227057546377182, + 0.0531110055744648, + 0.06141817942261696, + 0.022651800885796547, + 0.07408913969993591, + 0.12207674235105515, + 0.04689998924732208, + 0.040739189833402634, + 0.4415682256221771, + 0.04517512023448944 + ], + "credit_map": [ + 0.0173043105751276, + 0.013785187155008316, + 0.012368651106953621, + 0.012505736202001572, + 0.01099307183176279, + 0.007909195497632027, + 0.011733444407582283, + 0.005670985206961632, + 0.01343009527772665, + 0.008658490143716335, + 0.005351131781935692, + 0.006381002254784107, + 0.005039832554757595, + 0.0056420122273266315, + 0.003890424966812134, + 0.004718021024018526, + 0.008149994537234306, + 0.00635804608464241, + 0.00816104095429182, + 0.007978747598826885, + 0.007476205006241798, + 0.009321026504039764, + 0.007929875515401363, + 0.00604325020685792, + 0.004879183601588011, + 0.0031644199043512344, + 0.003220152575522661, + 0.0017759171314537525, + 0.002666594460606575, + 0.0028492319397628307, + 0.0024169536773115396, + 0.001679347944445908, + 0.009764597751200199, + 0.005563342012465, + 0.006596600636839867, + 0.005503714084625244, + 0.007007240783423185, + 0.01564483530819416, + 0.018434107303619385, + 0.005574701353907585, + 0.021619487553834915, + 0.019448861479759216, + 0.011393127962946892, + 0.011070851236581802, + 0.013385350815951824, + 0.014969303272664547, + 0.019813530147075653, + 0.010376225225627422, + 0.006863914895802736, + 0.005518076941370964, + 0.004970512352883816, + 0.004498308524489403, + 0.0065484982915222645, + 0.005732674617320299, + 0.005496626254171133, + 0.007271378301084042, + 0.007055611349642277, + 0.003999046515673399, + 0.005569301079958677, + 0.00562361441552639, + 0.005769607611000538, + 0.0037916037254035473, + 0.003564495826140046, + 0.005365910939872265, + 0.03919654339551926, + 0.051140546798706055, + 0.0869860127568245, + 0.12218891084194183, + 0.04638092219829559, + 0.01779639720916748, + 0.04552498832345009, + 0.03235393390059471, + 0.006832909304648638, + 0.0072480845265090466, + 0.009340987540781498, + 0.008815665729343891, + 0.00505139259621501, + 0.001960763242095709, + 0.0030528451316058636, + 0.0028724721632897854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.1798684448003769 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, rendered in the stark, beautifully minimalist style of Chinese ink painting, begins with a slow, deliberate downward tilt through a dense jungle of short, wide bamboo. The entire scene is a masterful performance of monochrome tones against a textured background of grayish-white rice paper, with soft, ethereal snowflakes silently drifting down. As the perspective descends, it reveals a subtle contrast between the pure, untouched white of the snow and the dark, damp surface of the bamboo. The broad leaves are depicted with bold, fluid strokes of ink, a deep, saturated black, their surfaces shimmering where the snow has yet to fall. The segmented stalks, rendered in varying shades of gray, their strong vertical forms contrast with the soft, horizontal layers of snow. The slow camera movement induces a contemplative focus, drawing the eye to the subtle interactions of snow and vegetation—a leaf gently bending under the weight of fresh powder, the sharp line where white snow meets black stem—creating an atmosphere of profound tranquility and minimalist beauty.", + "seed": 4150, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_6431_seed_4150_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": -0.44913059994578364, + "vlm_logprob": -0.44913059994578364, + "vlm_yes_no_margin": 1.55, + "vlm_token_credit": -0.44913059994578364, + "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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5786861181259155 + ], + "opposite_logprob_tokens": [ + -0.9536861181259155 + ], + "logprob_reward": -0.5786861181259155, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5786861181259155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1052979975938797, + 0.17436479032039642, + 0.3101261854171753, + 0.09993976354598999, + 0.03227854520082474, + 0.06478093564510345, + 0.04975654557347298, + 0.060347579419612885, + 0.04885057732462883, + 0.05425700917840004 + ], + "credit_map": [ + 0.016856590285897255, + 0.011673524044454098, + 0.011134544387459755, + 0.01203468069434166, + 0.009026910178363323, + 0.016214115545153618, + 0.011994749307632446, + 0.016362888738512993, + 0.022985033690929413, + 0.011159438639879227, + 0.011346593499183655, + 0.014801804907619953, + 0.00942709669470787, + 0.04527286812663078, + 0.046300701797008514, + 0.013071252033114433, + 0.02796885371208191, + 0.027332091704010963, + 0.03126247227191925, + 0.026261446997523308, + 0.08580835163593292, + 0.03905104100704193, + 0.036891594529151917, + 0.03555034473538399, + 0.008665502071380615, + 0.008598245680332184, + 0.007972318679094315, + 0.011921463534235954, + 0.017947610467672348, + 0.016000544652342796, + 0.015770800411701202, + 0.0130632808431983, + 0.004371337592601776, + 0.0050363922491669655, + 0.0051943897269666195, + 0.003969368990510702, + 0.0031509450636804104, + 0.0038359486497938633, + 0.0042617470026016235, + 0.002458415227010846, + 0.004537254571914673, + 0.004665466025471687, + 0.004534071311354637, + 0.00245769415050745, + 0.018440060317516327, + 0.00696974154561758, + 0.002967077773064375, + 0.020209571346640587, + 0.0058274767361581326, + 0.004265322349965572, + 0.008091883733868599, + 0.0046901339665055275, + 0.006342161446809769, + 0.008447748608887196, + 0.004974521230906248, + 0.007117298897355795, + 0.007905152626335621, + 0.010525541380047798, + 0.010879808105528355, + 0.00578455813229084, + 0.00575122470036149, + 0.006146920844912529, + 0.006093273870646954, + 0.0072611006908118725, + 0.004779416602104902, + 0.004221615847200155, + 0.0072134812362492085, + 0.005250339396297932, + 0.007172422483563423, + 0.007614009082317352, + 0.005741743836551905, + 0.006857545580714941, + 0.007641478907316923, + 0.007812920957803726, + 0.00810807291418314, + 0.005917955189943314, + 0.005947194993495941, + 0.005032731685787439, + 0.005754261743277311, + 0.008042393252253532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.5786861181259155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.502058744430542 + ], + "opposite_logprob_tokens": [ + -0.252058744430542 + ], + "logprob_reward": -1.502058744430542, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502058744430542, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16533972322940826, + 0.11647987365722656, + 0.14113861322402954, + 0.10619810223579407, + 0.07783026248216629, + 0.1275908648967743, + 0.10091787576675415, + 0.05945323035120964, + 0.06212351471185684, + 0.04292796179652214 + ], + "credit_map": [ + 0.02521735057234764, + 0.017223482951521873, + 0.014196998439729214, + 0.020534677430987358, + 0.013224538415670395, + 0.030207205563783646, + 0.027483100071549416, + 0.017252355813980103, + 0.022850703448057175, + 0.007426464930176735, + 0.009147312492132187, + 0.007211961317807436, + 0.006965494714677334, + 0.024906771257519722, + 0.026476765051484108, + 0.011494397185742855, + 0.013857120648026466, + 0.012474177405238152, + 0.014199663884937763, + 0.011509115807712078, + 0.03723713383078575, + 0.018283545970916748, + 0.01717235893011093, + 0.016405483707785606, + 0.008906804025173187, + 0.010856818407773972, + 0.010404209606349468, + 0.013946340419352055, + 0.025069573894143105, + 0.015202430076897144, + 0.012508385814726353, + 0.009303532540798187, + 0.005445452872663736, + 0.005675259046256542, + 0.007216853089630604, + 0.0039549218490719795, + 0.021423732861876488, + 0.007228601258248091, + 0.0067983586341142654, + 0.02008707821369171, + 0.008437289856374264, + 0.009907115250825882, + 0.0030677916947752237, + 0.002760150469839573, + 0.045364849269390106, + 0.008700307458639145, + 0.004095963202416897, + 0.04525739699602127, + 0.010941877961158752, + 0.01218888908624649, + 0.012191676534712315, + 0.00792077835649252, + 0.014314725995063782, + 0.019481124356389046, + 0.009421074762940407, + 0.014457731507718563, + 0.007830456830561161, + 0.010957160033285618, + 0.009930189698934555, + 0.004261092282831669, + 0.006680339574813843, + 0.006769080646336079, + 0.004591379780322313, + 0.008433524519205093, + 0.006589156575500965, + 0.005978631321340799, + 0.010728423483669758, + 0.006597326602786779, + 0.00918618869036436, + 0.007978775538504124, + 0.005475661717355251, + 0.009589348919689655, + 0.004378405399620533, + 0.0044876327738165855, + 0.007821430452167988, + 0.0039151920937001705, + 0.005459415260702372, + 0.004368211608380079, + 0.005500147584825754, + 0.00699752289801836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -1.502058744430542 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054994579404592514 + ], + "opposite_logprob_tokens": [ + -2.929994583129883 + ], + "logprob_reward": -0.054994579404592514, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054994579404592514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0931624323129654, + 0.3840298652648926, + 0.12137479335069656, + 0.06919153779745102, + 0.02800118364393711, + 0.17856323719024658, + 0.02150914818048477, + 0.03469790518283844, + 0.02386336959898472, + 0.04560650512576103 + ], + "credit_map": [ + 0.013090534135699272, + 0.008176474831998348, + 0.00944117084145546, + 0.007835565134882927, + 0.014938323758542538, + 0.012608787976205349, + 0.016727907583117485, + 0.010343668982386589, + 0.021756118163466454, + 0.008040163666009903, + 0.013616716489195824, + 0.02287384308874607, + 0.020886562764644623, + 0.13610941171646118, + 0.1352778673171997, + 0.025469209998846054, + 0.013274993747472763, + 0.01489592157304287, + 0.009519074112176895, + 0.011914371512830257, + 0.026888474822044373, + 0.018335573375225067, + 0.013113939203321934, + 0.013432459905743599, + 0.0061965142376720905, + 0.005562248174101114, + 0.005201575346291065, + 0.005568040069192648, + 0.015464802272617817, + 0.012903759256005287, + 0.009311514906585217, + 0.008983088657259941, + 0.0028603095561265945, + 0.0019351429073140025, + 0.0037030784878879786, + 0.001858231727965176, + 0.0031963034998625517, + 0.004870920442044735, + 0.004559083841741085, + 0.005018115509301424, + 0.01113020721822977, + 0.01196120772510767, + 0.005327735096216202, + 0.004061446525156498, + 0.06408803164958954, + 0.013803236186504364, + 0.006434462498873472, + 0.0617569163441658, + 0.0025511228013783693, + 0.0024148838128894567, + 0.003725918708369136, + 0.001972244121134281, + 0.0025124032981693745, + 0.003473107237368822, + 0.0021435199305415154, + 0.0027159505989402533, + 0.0034727740567177534, + 0.004743593279272318, + 0.006815189030021429, + 0.0031306184828281403, + 0.004008673131465912, + 0.004556853324174881, + 0.0032950721215456724, + 0.004675132688134909, + 0.0016387015348300338, + 0.0018314035842195153, + 0.0056903851218521595, + 0.0020426323171705008, + 0.004013971425592899, + 0.0022221507970243692, + 0.0025428631342947483, + 0.003881264477968216, + 0.006411913316696882, + 0.006313114892691374, + 0.007564539555460215, + 0.004176969174295664, + 0.006428957916796207, + 0.004368852358311415, + 0.004196168854832649, + 0.006145993247628212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.054994579404592514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054972801357507706 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.054972801357507706, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054972801357507706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07000639289617538, + 0.08710023760795593, + 0.05315243825316429, + 0.029016083106398582, + 0.02574264444410801, + 0.13449335098266602, + 0.034299470484256744, + 0.09745849668979645, + 0.08889926224946976, + 0.37983155250549316 + ], + "credit_map": [ + 0.010494589805603027, + 0.007127099204808474, + 0.009620721451938152, + 0.0064926850609481335, + 0.008584928698837757, + 0.008693632669746876, + 0.010703430511057377, + 0.008289310149848461, + 0.007405395619571209, + 0.003797238925471902, + 0.0056005134247243404, + 0.005907257553189993, + 0.004625855013728142, + 0.02297442965209484, + 0.023848699405789375, + 0.012940851040184498, + 0.004220794420689344, + 0.004142270423471928, + 0.0043036602437496185, + 0.004479810129851103, + 0.011674678884446621, + 0.009993834421038628, + 0.006474208552390337, + 0.007863184437155724, + 0.0027619944885373116, + 0.0022791095543652773, + 0.002874186495319009, + 0.0033987185452133417, + 0.00394746009260416, + 0.004728551022708416, + 0.004529053810983896, + 0.004497011657804251, + 0.0027495799586176872, + 0.002241449663415551, + 0.0034848973155021667, + 0.0026472522877156734, + 0.003048845799639821, + 0.004389169625937939, + 0.0031547846738249063, + 0.0040266672149300575, + 0.00903992634266615, + 0.008155087009072304, + 0.004161643795669079, + 0.00544964661821723, + 0.046623922884464264, + 0.00923230778425932, + 0.007543571759015322, + 0.04428727179765701, + 0.003476191544905305, + 0.0032096169888973236, + 0.004994803108274937, + 0.0034567848779261112, + 0.0052098119631409645, + 0.005671137943863869, + 0.0033085751347243786, + 0.004972550552338362, + 0.01015070453286171, + 0.012543845921754837, + 0.017288003116846085, + 0.007216386962682009, + 0.013293584808707237, + 0.01649503782391548, + 0.007820170372724533, + 0.012650765478610992, + 0.008496427908539772, + 0.008493575267493725, + 0.012544042430818081, + 0.008668319322168827, + 0.010901292786002159, + 0.013594047166407108, + 0.012706336565315723, + 0.013495225459337234, + 0.05057133361697197, + 0.03492557629942894, + 0.07851853966712952, + 0.03381193056702614, + 0.04665558785200119, + 0.04099377989768982, + 0.03437869995832443, + 0.05997613072395325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.054972801357507706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054940756410360336 + ], + "opposite_logprob_tokens": [ + -2.929940700531006 + ], + "logprob_reward": -0.054940756410360336, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054940756410360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05093211308121681, + 0.0714106410741806, + 0.053478457033634186, + 0.1505938619375229, + 0.03700142726302147, + 0.47775664925575256, + 0.029867704957723618, + 0.05045342072844505, + 0.0500531867146492, + 0.028452519327402115 + ], + "credit_map": [ + 0.008074619807302952, + 0.006259525660425425, + 0.005972079001367092, + 0.004835512954741716, + 0.005465191788971424, + 0.0068870107643306255, + 0.006274189800024033, + 0.0071639870293438435, + 0.006030323449522257, + 0.004575887694954872, + 0.005756632424890995, + 0.004072895273566246, + 0.008988630957901478, + 0.015788566321134567, + 0.01459356490522623, + 0.011604144237935543, + 0.005794194061309099, + 0.004805833101272583, + 0.0061278208158910275, + 0.004464018624275923, + 0.013087027706205845, + 0.0064651332795619965, + 0.006392493378371, + 0.0063419402576982975, + 0.012540237978100777, + 0.012453130446374416, + 0.012187755666673183, + 0.015079299919307232, + 0.036033306270837784, + 0.02721966803073883, + 0.01905476115643978, + 0.016025710850954056, + 0.003452010452747345, + 0.00385511782951653, + 0.005184977315366268, + 0.00330159324221313, + 0.006476313807070255, + 0.004259761422872543, + 0.004683074541389942, + 0.005788580980151892, + 0.030147766694426537, + 0.03266455978155136, + 0.016302283853292465, + 0.01061514113098383, + 0.17206348478794098, + 0.03440938889980316, + 0.017264267429709435, + 0.16428978741168976, + 0.002859037136659026, + 0.0035678513813763857, + 0.005169074982404709, + 0.002414063084870577, + 0.0047262986190617085, + 0.0037978903856128454, + 0.0028242897242307663, + 0.004509201738983393, + 0.006506224162876606, + 0.009822499006986618, + 0.00772497383877635, + 0.003848989261314273, + 0.005093758925795555, + 0.006708830129355192, + 0.0036848648451268673, + 0.007063285913318396, + 0.004400181118398905, + 0.005106697790324688, + 0.00804821401834488, + 0.005440392531454563, + 0.00535053014755249, + 0.00842614658176899, + 0.005554682575166225, + 0.007726343814283609, + 0.0037009655497968197, + 0.0026717078872025013, + 0.006417023949325085, + 0.0023439479991793633, + 0.0032782440539449453, + 0.003256997326388955, + 0.0020523816347122192, + 0.004731253255158663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.054940756410360336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, rendered in the stark, beautifully minimalist style of Chinese ink painting, begins with a slow, deliberate downward tilt through a dense jungle of short, wide bamboo. The entire scene is a masterful performance of monochrome tones against a textured background of grayish-white rice paper, with soft, ethereal snowflakes silently drifting down. As the perspective descends, it reveals a subtle contrast between the pure, untouched white of the snow and the dark, damp surface of the bamboo. The broad leaves are depicted with bold, fluid strokes of ink, a deep, saturated black, their surfaces shimmering where the snow has yet to fall. The segmented stalks, rendered in varying shades of gray, their strong vertical forms contrast with the soft, horizontal layers of snow. The slow camera movement induces a contemplative focus, drawing the eye to the subtle interactions of snow and vegetation—a leaf gently bending under the weight of fresh powder, the sharp line where white snow meets black stem—creating an atmosphere of profound tranquility and minimalist beauty.", + "seed": 4151, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_6431_seed_4151_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": -0.2980147484689951, + "vlm_logprob": -0.2980147484689951, + "vlm_yes_no_margin": 1.9250000238418579, + "vlm_token_credit": -0.2980147484689951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6186815500259399 + ], + "opposite_logprob_tokens": [ + -0.8686815500259399 + ], + "logprob_reward": -0.6186815500259399, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.6186815500259399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09932637214660645, + 0.09572072327136993, + 0.15883895754814148, + 0.1698426455259323, + 0.09414432942867279, + 0.04819023236632347, + 0.06614270061254501, + 0.09873327612876892, + 0.09341545403003693, + 0.07564527541399002 + ], + "credit_map": [ + 0.012760092504322529, + 0.030551398172974586, + 0.00585097586736083, + 0.00932772271335125, + 0.009588590823113918, + 0.016716254875063896, + 0.005890841130167246, + 0.008640498854219913, + 0.020806748420000076, + 0.012204625643789768, + 0.007941345684230328, + 0.0068956357426941395, + 0.01876830868422985, + 0.012528441846370697, + 0.006612120661884546, + 0.00996349286288023, + 0.006711429916322231, + 0.017062747851014137, + 0.008193528279662132, + 0.010494777001440525, + 0.018242983147501945, + 0.0517098605632782, + 0.008914406411349773, + 0.037509236484766006, + 0.014018839225172997, + 0.040293559432029724, + 0.01213556807488203, + 0.02011793665587902, + 0.015872303396463394, + 0.03927350789308548, + 0.014716176316142082, + 0.01341476384550333, + 0.012340810149908066, + 0.01248814258724451, + 0.010577374137938023, + 0.012137130834162235, + 0.011404644697904587, + 0.016199547797441483, + 0.009180819615721703, + 0.009815861470997334, + 0.0063414196483790874, + 0.005578568670898676, + 0.004567072726786137, + 0.005763125605881214, + 0.0050504738464951515, + 0.006835194304585457, + 0.006940396036952734, + 0.00711398059502244, + 0.00560512812808156, + 0.013825464993715286, + 0.006406629923731089, + 0.00586942583322525, + 0.011022438295185566, + 0.007552310824394226, + 0.009038449265062809, + 0.00682285288348794, + 0.008133151568472385, + 0.01553019043058157, + 0.01056874729692936, + 0.008743754588067532, + 0.014719929546117783, + 0.014439414255321026, + 0.011993263848125935, + 0.014604817144572735, + 0.006104073952883482, + 0.0168670155107975, + 0.006644000764936209, + 0.0072823213413357735, + 0.01148341316729784, + 0.021956535056233406, + 0.012551169842481613, + 0.010526929050683975, + 0.0066201644949615, + 0.005287588108330965, + 0.004722703713923693, + 0.006261981092393398, + 0.017431532964110374, + 0.00645408034324646, + 0.008036439307034016, + 0.02083078771829605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.6186815500259399 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327987313270569 + ], + "opposite_logprob_tokens": [ + -0.7577987313270569 + ], + "logprob_reward": -0.6327987313270569, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327987313270569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10147659480571747, + 0.11203824728727341, + 0.09666594117879868, + 0.10418230295181274, + 0.0679386705160141, + 0.058452121913433075, + 0.14439335465431213, + 0.08743397146463394, + 0.08337131887674332, + 0.14404745399951935 + ], + "credit_map": [ + 0.0073057981207966805, + 0.02663707360625267, + 0.007675196509808302, + 0.010094263590872288, + 0.011485233902931213, + 0.01675524190068245, + 0.00856990460306406, + 0.012953884899616241, + 0.007029738277196884, + 0.01949559897184372, + 0.005595263559371233, + 0.008095037192106247, + 0.03504164516925812, + 0.01760728470981121, + 0.011302306316792965, + 0.007871379144489765, + 0.0028293689247220755, + 0.008371737785637379, + 0.00362945138476789, + 0.004636019002646208, + 0.005719424691051245, + 0.03654322773218155, + 0.003671797225251794, + 0.03126491233706474, + 0.010265166871249676, + 0.016974693164229393, + 0.005800600163638592, + 0.00953298807144165, + 0.006066258531063795, + 0.025572240352630615, + 0.007616591174155474, + 0.02235376462340355, + 0.005906798876821995, + 0.0075657665729522705, + 0.008457083255052567, + 0.008965687826275826, + 0.007985750213265419, + 0.012641817331314087, + 0.006800976116210222, + 0.009614787995815277, + 0.00603785365819931, + 0.007826903834939003, + 0.007114037871360779, + 0.009368231520056725, + 0.005923423916101456, + 0.007482893764972687, + 0.0065796407870948315, + 0.008119137957692146, + 0.02187461219727993, + 0.022722195833921432, + 0.013352039270102978, + 0.010148199275135994, + 0.026334373280405998, + 0.01927844062447548, + 0.014299271628260612, + 0.016384227201342583, + 0.008549109101295471, + 0.013207094743847847, + 0.00770598417147994, + 0.008129272609949112, + 0.012059521861374378, + 0.011188619770109653, + 0.008611006662249565, + 0.017983360216021538, + 0.006896044593304396, + 0.014206431806087494, + 0.010161182843148708, + 0.008296096697449684, + 0.010784879326820374, + 0.012243668548762798, + 0.013537751510739326, + 0.007245260290801525, + 0.013625241816043854, + 0.011558582074940205, + 0.01031116396188736, + 0.019131042063236237, + 0.02426125667989254, + 0.028646033257246017, + 0.011785920709371567, + 0.02472822181880474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.6327987313270569 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754660978913307 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.006754660978913307, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754660978913307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15851210057735443, + 0.1160973310470581, + 0.12655994296073914, + 0.04185895249247551, + 0.05501818284392357, + 0.03302573785185814, + 0.3090071678161621, + 0.049162186682224274, + 0.028580131009221077, + 0.0821782574057579 + ], + "credit_map": [ + 0.015436701476573944, + 0.03329513594508171, + 0.0128654595464468, + 0.02177719585597515, + 0.016449304297566414, + 0.026784207671880722, + 0.014848846942186356, + 0.017055250704288483, + 0.01993413269519806, + 0.014713305979967117, + 0.011224251240491867, + 0.011766994372010231, + 0.01960926502943039, + 0.013952763751149178, + 0.011138727888464928, + 0.013757900334894657, + 0.0031440213788300753, + 0.011010760441422462, + 0.0038870994467288256, + 0.004825668875128031, + 0.0061059389263391495, + 0.049902383238077164, + 0.0037218842189759016, + 0.04396218806505203, + 0.004105876665562391, + 0.009410957805812359, + 0.0029742009937763214, + 0.004856038372963667, + 0.0037934412248432636, + 0.009090282954275608, + 0.0036993902176618576, + 0.003928765654563904, + 0.005550084635615349, + 0.007778409868478775, + 0.006810570601373911, + 0.007246366702020168, + 0.0068431454710662365, + 0.010106027126312256, + 0.004729076288640499, + 0.0059545026160776615, + 0.0032187197357416153, + 0.0038506430573761463, + 0.003022742923349142, + 0.007292871363461018, + 0.004360254388302565, + 0.002888928633183241, + 0.003772821044549346, + 0.00461875693872571, + 0.03786693513393402, + 0.05648051202297211, + 0.023985082283616066, + 0.020537665113806725, + 0.07921722531318665, + 0.03282196819782257, + 0.02747196890413761, + 0.03062581643462181, + 0.004083868116140366, + 0.006245429161936045, + 0.004242230206727982, + 0.00494701974093914, + 0.005339347757399082, + 0.005871477536857128, + 0.007548652123659849, + 0.010884164832532406, + 0.002073949435725808, + 0.005708358716219664, + 0.003008901607245207, + 0.0025963729713112116, + 0.00369578436948359, + 0.00556519441306591, + 0.0032479632645845413, + 0.002683605533093214, + 0.007593564223498106, + 0.007461024448275566, + 0.005018857307732105, + 0.006292846519500017, + 0.01800847239792347, + 0.007559395395219326, + 0.008697024546563625, + 0.021547066047787666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.006754660978913307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282453060150146 + ], + "opposite_logprob_tokens": [ + -2.017824649810791 + ], + "logprob_reward": -0.14282453060150146, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282453060150146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15775227546691895, + 0.1233472004532814, + 0.1915353387594223, + 0.10252148658037186, + 0.08904938399791718, + 0.07842347770929337, + 0.07171033322811127, + 0.05643383413553238, + 0.06002595275640488, + 0.06920075416564941 + ], + "credit_map": [ + 0.013909630477428436, + 0.035363033413887024, + 0.011828956194221973, + 0.02695511095225811, + 0.015745054930448532, + 0.028014162555336952, + 0.013191495090723038, + 0.012744824402034283, + 0.014614447019994259, + 0.014394759200513363, + 0.009044854901731014, + 0.012692341580986977, + 0.03241736441850662, + 0.014926384203135967, + 0.012512526474893093, + 0.012744520790874958, + 0.013662446290254593, + 0.020889803767204285, + 0.01070066075772047, + 0.01406219881027937, + 0.021086355671286583, + 0.04918120428919792, + 0.011906604282557964, + 0.05004604905843735, + 0.00976339541375637, + 0.018405139446258545, + 0.006264673080295324, + 0.012017663568258286, + 0.007800746243447065, + 0.022516749799251556, + 0.012600325979292393, + 0.013152789324522018, + 0.009047262370586395, + 0.010496015660464764, + 0.010869095101952553, + 0.009644107893109322, + 0.011070085689425468, + 0.015862390398979187, + 0.008671221323311329, + 0.013389206491410732, + 0.006252739578485489, + 0.008486036211252213, + 0.007071361877024174, + 0.009785871021449566, + 0.006880538538098335, + 0.009774900041520596, + 0.006523917894810438, + 0.023648109287023544, + 0.011696940287947655, + 0.009124242700636387, + 0.006634482182562351, + 0.006353872362524271, + 0.010303529910743237, + 0.010585397481918335, + 0.008208993822336197, + 0.008802869357168674, + 0.005421817302703857, + 0.008340948261320591, + 0.0052308738231658936, + 0.005131589714437723, + 0.008660140447318554, + 0.007514652330428362, + 0.006100738886743784, + 0.010033069178462029, + 0.004302582703530788, + 0.008209358900785446, + 0.005049232859164476, + 0.005759773775935173, + 0.006921084132045507, + 0.010179092176258564, + 0.013390596024692059, + 0.0062142303213477135, + 0.00563817098736763, + 0.005795584991574287, + 0.0036700915079563856, + 0.0037732888013124466, + 0.021274667233228683, + 0.007076552603393793, + 0.004734246525913477, + 0.01723814569413662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.14282453060150146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901426941156387 + ], + "opposite_logprob_tokens": [ + -2.4640142917633057 + ], + "logprob_reward": -0.08901426941156387, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901426941156387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08311370760202408, + 0.11407940834760666, + 0.14319168031215668, + 0.06734278798103333, + 0.1815756857395172, + 0.11684801429510117, + 0.15376318991184235, + 0.1041347086429596, + 0.021041644737124443, + 0.014909191057085991 + ], + "credit_map": [ + 0.008357709273695946, + 0.017912862822413445, + 0.006263969466090202, + 0.0063285850919783115, + 0.011200531385838985, + 0.0136656379327178, + 0.007839633151888847, + 0.01154477521777153, + 0.007523882668465376, + 0.01704264245927334, + 0.007370657287538052, + 0.009926687926054, + 0.03224259242415428, + 0.015894025564193726, + 0.013655981048941612, + 0.010422945953905582, + 0.0033403881825506687, + 0.011862745508551598, + 0.004417534451931715, + 0.005589183419942856, + 0.0076445043087005615, + 0.05708729475736618, + 0.0039057908579707146, + 0.049344222992658615, + 0.006285463459789753, + 0.014491549693048, + 0.00463003059849143, + 0.008228139951825142, + 0.0067682391963899136, + 0.01003832183778286, + 0.006443066988140345, + 0.01045797299593687, + 0.01641726866364479, + 0.020127160474658012, + 0.01877923123538494, + 0.024996140971779823, + 0.02281022071838379, + 0.03432527184486389, + 0.01919107511639595, + 0.02492930367588997, + 0.0128583163022995, + 0.013713966123759747, + 0.013760283589363098, + 0.0191948264837265, + 0.01100841723382473, + 0.01608404330909252, + 0.013714674860239029, + 0.016513487324118614, + 0.019661923870444298, + 0.027599286288022995, + 0.017436692491173744, + 0.013593567535281181, + 0.017478523775935173, + 0.016408778727054596, + 0.02120134048163891, + 0.0203830786049366, + 0.00914568267762661, + 0.014628596603870392, + 0.007185438182204962, + 0.007441414520144463, + 0.018282316625118256, + 0.013878988102078438, + 0.017179230228066444, + 0.016393041238188744, + 0.001308670500293374, + 0.003412821562960744, + 0.0015164901269599795, + 0.0014528620522469282, + 0.002753326203674078, + 0.005943454802036285, + 0.002172259148210287, + 0.0024817604571580887, + 0.0010260603157803416, + 0.0009382439311593771, + 0.000669958651997149, + 0.0010486111277714372, + 0.004507202655076981, + 0.001926244469359517, + 0.0010432907147333026, + 0.0037495801225304604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.08901426941156387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, rendered in the stark, beautifully minimalist style of Chinese ink painting, begins with a slow, deliberate downward tilt through a dense jungle of short, wide bamboo. The entire scene is a masterful performance of monochrome tones against a textured background of grayish-white rice paper, with soft, ethereal snowflakes silently drifting down. As the perspective descends, it reveals a subtle contrast between the pure, untouched white of the snow and the dark, damp surface of the bamboo. The broad leaves are depicted with bold, fluid strokes of ink, a deep, saturated black, their surfaces shimmering where the snow has yet to fall. The segmented stalks, rendered in varying shades of gray, their strong vertical forms contrast with the soft, horizontal layers of snow. The slow camera movement induces a contemplative focus, drawing the eye to the subtle interactions of snow and vegetation—a leaf gently bending under the weight of fresh powder, the sharp line where white snow meets black stem—creating an atmosphere of profound tranquility and minimalist beauty.", + "seed": 4152, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_2/index_6431_seed_4152_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": -0.290597190707922, + "vlm_logprob": -0.290597190707922, + "vlm_yes_no_margin": 1.8, + "vlm_token_credit": -0.290597190707922, + "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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8178203701972961 + ], + "opposite_logprob_tokens": [ + -0.6928203701972961 + ], + "logprob_reward": -0.8178203701972961, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.8178203701972961, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14769360423088074, + 0.058872777968645096, + 0.1373351812362671, + 0.16440445184707642, + 0.2640397250652313, + 0.03705733269453049, + 0.04953346773982048, + 0.04218931868672371, + 0.02039858140051365, + 0.07847559452056885 + ], + "credit_map": [ + 0.014462369494140148, + 0.022218728438019753, + 0.033949632197618484, + 0.017163556069135666, + 0.010571246035397053, + 0.013415602035820484, + 0.01383925974369049, + 0.022073186933994293, + 0.00845794565975666, + 0.00846781488507986, + 0.005435746628791094, + 0.005257027689367533, + 0.008827614597976208, + 0.010466163977980614, + 0.004787816200405359, + 0.0071726455353200436, + 0.043129999190568924, + 0.04630325362086296, + 0.008963414467871189, + 0.004826662130653858, + 0.007754507940262556, + 0.009017021395266056, + 0.007127972785383463, + 0.010212341323494911, + 0.03270423412322998, + 0.022056905552744865, + 0.018674960359930992, + 0.01253278087824583, + 0.026251187548041344, + 0.020911987870931625, + 0.019170228391885757, + 0.012102152220904827, + 0.02655469998717308, + 0.01045480277389288, + 0.11370936781167984, + 0.025823744013905525, + 0.012781192548573017, + 0.02476639859378338, + 0.03559238836169243, + 0.014357132837176323, + 0.0023347078822553158, + 0.002683002268895507, + 0.009824056178331375, + 0.006144232582300901, + 0.004981926176697016, + 0.0024561267346143723, + 0.0037417106796056032, + 0.004891567397862673, + 0.005781720858067274, + 0.011001357808709145, + 0.004393185488879681, + 0.007525421679019928, + 0.009367931634187698, + 0.003347618505358696, + 0.004617030266672373, + 0.0034991963766515255, + 0.005774928722530603, + 0.008795497938990593, + 0.007045534905046225, + 0.00524152722209692, + 0.00654041999951005, + 0.0026058563962578773, + 0.0033212746493518353, + 0.0028642769902944565, + 0.0025863121263682842, + 0.0031019009184092283, + 0.0021535668056458235, + 0.0022786648478358984, + 0.0032408051192760468, + 0.0029573224019259214, + 0.0018370902398601174, + 0.0022429160308092833, + 0.02628987468779087, + 0.01377084944397211, + 0.01344367302954197, + 0.003559418022632599, + 0.00879041850566864, + 0.004501276649534702, + 0.003012074390426278, + 0.005107998847961426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.8178203701972961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42890438437461853 + ], + "opposite_logprob_tokens": [ + -1.053904414176941 + ], + "logprob_reward": -0.42890438437461853, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42890438437461853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07037893682718277, + 0.04423687234520912, + 0.07939936965703964, + 0.045444268733263016, + 0.04582871124148369, + 0.027330191805958748, + 0.04013756662607193, + 0.060692355036735535, + 0.0627322718501091, + 0.5238194465637207 + ], + "credit_map": [ + 0.005971122998744249, + 0.009335538372397423, + 0.014292769134044647, + 0.006642271298915148, + 0.0055707478895783424, + 0.009464790113270283, + 0.007650918327271938, + 0.011450777761638165, + 0.005854420363903046, + 0.005247063469141722, + 0.00590765243396163, + 0.005381740629673004, + 0.004457395989447832, + 0.007256071083247662, + 0.004215506371110678, + 0.0059170229360461235, + 0.029915830120444298, + 0.03157311677932739, + 0.0020376001484692097, + 0.001857774332165718, + 0.0038167256861925125, + 0.0031179420184344053, + 0.0034023637417703867, + 0.003678014734759927, + 0.006145095452666283, + 0.0051234206184744835, + 0.0043869055807590485, + 0.0035027589183300734, + 0.008565392345190048, + 0.008344707079231739, + 0.004958873614668846, + 0.004417116753757, + 0.0040627350099384785, + 0.00274290912784636, + 0.015922408550977707, + 0.004683175589889288, + 0.0051101199351251125, + 0.004725366365164518, + 0.005652989260852337, + 0.0029290097299963236, + 0.0022936766035854816, + 0.002235628431662917, + 0.004403519909828901, + 0.004964344669133425, + 0.00472188089042902, + 0.001862957957200706, + 0.0028811481315642595, + 0.00396703602746129, + 0.00421381089836359, + 0.007192622404545546, + 0.004798194393515587, + 0.004825292155146599, + 0.005630351137369871, + 0.005809267517179251, + 0.004205326084047556, + 0.0034626999404281378, + 0.005488959606736898, + 0.010683273896574974, + 0.006476461887359619, + 0.009016918018460274, + 0.01305440440773964, + 0.005518287420272827, + 0.0055120475590229034, + 0.00494200037792325, + 0.006445898208767176, + 0.00979661662131548, + 0.006832502316683531, + 0.005981067661195993, + 0.009718921035528183, + 0.0104091577231884, + 0.007292714901268482, + 0.006255393847823143, + 0.20138069987297058, + 0.09412311017513275, + 0.10751553624868393, + 0.023595495149493217, + 0.035510338842868805, + 0.014797086827456951, + 0.014195652678608894, + 0.03270150348544121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.42890438437461853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331206575036049 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331206575036049, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331206575036049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11117605119943619, + 0.06398266553878784, + 0.08548728376626968, + 0.050450798124074936, + 0.04820584878325462, + 0.032816868275403976, + 0.08582554012537003, + 0.27123335003852844, + 0.05106768012046814, + 0.19975382089614868 + ], + "credit_map": [ + 0.016063064336776733, + 0.02080361358821392, + 0.029075421392917633, + 0.010122861713171005, + 0.009820622391998768, + 0.010174846276640892, + 0.007291078567504883, + 0.007824541069567204, + 0.010617976076900959, + 0.007992204278707504, + 0.008463246747851372, + 0.007290450390428305, + 0.006946237292140722, + 0.010122687555849552, + 0.00502777798101306, + 0.007522080559283495, + 0.02213931642472744, + 0.022043297067284584, + 0.00554773909971118, + 0.004003899171948433, + 0.0058218189515173435, + 0.007589125074446201, + 0.008605902083218098, + 0.00973618496209383, + 0.00748547213152051, + 0.005943099968135357, + 0.004636418540030718, + 0.004810578189790249, + 0.00937056727707386, + 0.0076383184641599655, + 0.005704356823116541, + 0.004861985798925161, + 0.004804798401892185, + 0.004469884093850851, + 0.006132023874670267, + 0.006381644401699305, + 0.008778910152614117, + 0.0068346536718308926, + 0.005728989373892546, + 0.005074943415820599, + 0.002310829469934106, + 0.0020159694831818342, + 0.0067794593051075935, + 0.004836559761315584, + 0.00519799068570137, + 0.0029779821634292603, + 0.003617287613451481, + 0.00508078932762146, + 0.008243121206760406, + 0.02516762726008892, + 0.00934742670506239, + 0.008828760124742985, + 0.017838839441537857, + 0.006029571406543255, + 0.0061819711700081825, + 0.004188218154013157, + 0.03884022682905197, + 0.06512750685214996, + 0.04434589669108391, + 0.0320068895816803, + 0.04235471785068512, + 0.017220208421349525, + 0.014474059455096722, + 0.016863858327269554, + 0.006467862520366907, + 0.007249678019434214, + 0.00776192219927907, + 0.0050553930923342705, + 0.007136812899261713, + 0.006558127235621214, + 0.00630040280520916, + 0.004537481348961592, + 0.07300566881895065, + 0.036599040031433105, + 0.03901293873786926, + 0.010567544028162956, + 0.015682362020015717, + 0.005845334380865097, + 0.005800145212560892, + 0.013240806758403778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.02331206575036049 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700390413403511 + ], + "opposite_logprob_tokens": [ + -2.6950390338897705 + ], + "logprob_reward": -0.0700390413403511, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700390413403511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17941173911094666, + 0.21676424145698547, + 0.12464912235736847, + 0.04424114152789116, + 0.07138656824827194, + 0.022831596434116364, + 0.048557836562395096, + 0.04166368022561073, + 0.028874440118670464, + 0.2216196209192276 + ], + "credit_map": [ + 0.030774857848882675, + 0.035929299890995026, + 0.04372376576066017, + 0.01617007702589035, + 0.011438773944973946, + 0.014909429475665092, + 0.010994271375238895, + 0.01547124981880188, + 0.034553393721580505, + 0.025840671733021736, + 0.027216771617531776, + 0.025927672162652016, + 0.025140725076198578, + 0.03282686322927475, + 0.016998928040266037, + 0.02825920470058918, + 0.03419817239046097, + 0.03474295511841774, + 0.005262069404125214, + 0.00679128710180521, + 0.009213276207447052, + 0.010154811665415764, + 0.013032205402851105, + 0.01125434972345829, + 0.007683990057557821, + 0.0054741851054131985, + 0.004605826456099749, + 0.004597010090947151, + 0.007761550135910511, + 0.006465152837336063, + 0.0048896861262619495, + 0.0027637421153485775, + 0.006377162877470255, + 0.0039012013003230095, + 0.02678985707461834, + 0.008877660147845745, + 0.005399150308221579, + 0.0071618505753576756, + 0.00917050801217556, + 0.0037091763224452734, + 0.001759336912073195, + 0.002071884460747242, + 0.00402481947094202, + 0.005103036295622587, + 0.0021819632966071367, + 0.001539174816571176, + 0.003438084153458476, + 0.0027132986579090357, + 0.004607925657182932, + 0.005893950816243887, + 0.007402024231851101, + 0.00713831651955843, + 0.0065598320215940475, + 0.005767082329839468, + 0.006177513860166073, + 0.00501119252294302, + 0.006203218828886747, + 0.006283419206738472, + 0.0063634468242526054, + 0.004633873235434294, + 0.006529057864099741, + 0.0032450808212161064, + 0.005616768263280392, + 0.002788818208500743, + 0.0034755149390548468, + 0.004614416044205427, + 0.0031361414585262537, + 0.0035465836990624666, + 0.0047810436226427555, + 0.0037848055362701416, + 0.0033217635937035084, + 0.0022141719236969948, + 0.08586902916431427, + 0.03990165889263153, + 0.04488703981041908, + 0.00995536521077156, + 0.015265737660229206, + 0.006107657216489315, + 0.005983651150017977, + 0.013649484142661095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0700390413403511 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 as a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a dense jungle of short, wide bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style rendered in a stark, minimalist Chinese ink painting style on a grayish-white rice paper background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soft, ethereal snowflakes visible drifting down in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow, deliberate downward tilt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291009187698364 + ], + "opposite_logprob_tokens": [ + -2.237910032272339 + ], + "logprob_reward": -0.11291009187698364, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291009187698364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1326669603586197, + 0.04561352729797363, + 0.046290021389722824, + 0.13622914254665375, + 0.07565625756978989, + 0.040376417338848114, + 0.03379620611667633, + 0.025677813217043877, + 0.02550465799868107, + 0.43818899989128113 + ], + "credit_map": [ + 0.016582289710640907, + 0.020269237458705902, + 0.03479110449552536, + 0.01305447518825531, + 0.010406163521111012, + 0.01362405065447092, + 0.009272899478673935, + 0.014666732400655746, + 0.005689093843102455, + 0.006334715988487005, + 0.0035237777046859264, + 0.004338232334703207, + 0.007804746273905039, + 0.00831630453467369, + 0.003673696657642722, + 0.005932962987571955, + 0.010212966240942478, + 0.010902183130383492, + 0.004717651754617691, + 0.004558559041470289, + 0.0033079427666962147, + 0.004518153145909309, + 0.004429807886481285, + 0.0036427576560527086, + 0.02709764987230301, + 0.017859121784567833, + 0.015482485294342041, + 0.012708858586847782, + 0.019931098446249962, + 0.01630220375955105, + 0.01625552400946617, + 0.010592195205390453, + 0.006502963602542877, + 0.003958842717111111, + 0.030853524804115295, + 0.00929169449955225, + 0.00439697178080678, + 0.0074104489758610725, + 0.00941385980695486, + 0.0038279497530311346, + 0.0031274540815502405, + 0.003547310596331954, + 0.005492561496794224, + 0.007942702621221542, + 0.006043091882020235, + 0.0031344795133918524, + 0.004954906180500984, + 0.006133911665529013, + 0.003317448543384671, + 0.007048172876238823, + 0.004793253727257252, + 0.004387971479445696, + 0.0059671662747859955, + 0.0029642346780747175, + 0.003169459756463766, + 0.0021484973840415478, + 0.004064771346747875, + 0.004150567576289177, + 0.004756055306643248, + 0.0022036447189748287, + 0.003805747488513589, + 0.0017422272358089685, + 0.0026227396447211504, + 0.0023320589680224657, + 0.0028118202462792397, + 0.0034749163314700127, + 0.0025492722634226084, + 0.0028098071925342083, + 0.0047447700053453445, + 0.0037058107554912567, + 0.0030218889005482197, + 0.002386372769251466, + 0.16492319107055664, + 0.08141212910413742, + 0.08621487766504288, + 0.0195603184401989, + 0.03232152760028839, + 0.012836801819503307, + 0.012026350013911724, + 0.02889380417764187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.11291009187698364 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_tensor([3024, 3024])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_tensor([3024, 3024])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3e77e1ee8aea0bade154e143455613da7afbce02 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_tensor([3024, 3024])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Low-angle, static medium shots capture the moment of violent impact in breathtaking super-slow motion, rendered in high-quality cinematic detail. On a damp, bumpy cobblestone street, shrouded in mystery, a fantastical, fuel-powered tricycle collides with an unseen force. The vehicle, a masterpiece of ornate Victorian design, its frame adorned with polished brass fittings and intricate gears, gleams in the dim light. As the front of the tricycle violently collapses, the crimson velvet seats ripple with the shockwave. The true force of the impact is visualized as a nearby cast-iron gas lamp shatters, sending a cascade of glittering glass shards streaking gracefully through the air in mesmerizing slow motion. These fragments, capturing the flickering warm golden glow of the dying gas flame, seem suspended in thick diamonds as the swirling fog swirls and separates in chaos, creating a scene of intense and haunting beauty, freezing a single, devastating moment.", + "seed": 436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_3024_seed_436_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": -0.06259189704433084, + "vlm_logprob": -0.06259189704433084, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.06259189704433084, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.17964395880699158 + ], + "opposite_logprob_tokens": [ + -2.8046438694000244 + ], + "logprob_reward": -0.17964395880699158, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.17964395880699158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1541597992181778, + 0.08850167691707611, + 0.03656885027885437, + 0.36751025915145874, + 0.033822350203990936, + 0.05345138907432556, + 0.08160070329904556, + 0.07613632082939148, + 0.058267805725336075, + 0.049980852752923965 + ], + "credit_map": [ + 0.01566685549914837, + 0.020101645961403847, + 0.010934078134596348, + 0.013133984990417957, + 0.015310565009713173, + 0.03735070303082466, + 0.024731524288654327, + 0.01693044975399971, + 0.007950970903038979, + 0.012540653347969055, + 0.014609618112444878, + 0.005821797996759415, + 0.00873293075710535, + 0.024415189400315285, + 0.008272231556475163, + 0.006158282049000263, + 0.0057351295836269855, + 0.005828912369906902, + 0.005461276043206453, + 0.0021083729807287455, + 0.003428623080253601, + 0.005558470729738474, + 0.0035923549439758062, + 0.004855715204030275, + 0.025268159806728363, + 0.0182940810918808, + 0.1155492439866066, + 0.02878049574792385, + 0.019481763243675232, + 0.04163019359111786, + 0.01523304171860218, + 0.10327330231666565, + 0.00768626993522048, + 0.003996293991804123, + 0.003506541484966874, + 0.003931804560124874, + 0.00346927042119205, + 0.003073501167818904, + 0.0029802548233419657, + 0.005178411956876516, + 0.012735594063997269, + 0.007168677169829607, + 0.005362569820135832, + 0.004291974939405918, + 0.007270174566656351, + 0.005435530561953783, + 0.006455540657043457, + 0.00473132636398077, + 0.013318201526999474, + 0.008573084138333797, + 0.012296236120164394, + 0.01595638319849968, + 0.007839091122150421, + 0.008694569580256939, + 0.007666460238397121, + 0.00725668016821146, + 0.010550404898822308, + 0.014172904193401337, + 0.007966239005327225, + 0.0155165521427989, + 0.005020757205784321, + 0.009190136566758156, + 0.007927663624286652, + 0.00579166179522872, + 0.006546216085553169, + 0.011137169785797596, + 0.015893397852778435, + 0.008143470622599125, + 0.004128595348447561, + 0.0033782441169023514, + 0.005478277336806059, + 0.003562433645129204, + 0.007994384504854679, + 0.004481703974306583, + 0.007382803130894899, + 0.010559403337538242, + 0.004718575160950422, + 0.004389688838273287, + 0.0050569986924529076, + 0.005397295579314232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.17964395880699158 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.0337490439414978 + ], + "opposite_logprob_tokens": [ + -3.4087491035461426 + ], + "logprob_reward": -0.0337490439414978, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337490439414978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1891096979379654, + 0.032030317932367325, + 0.14591079950332642, + 0.15444684028625488, + 0.01271491963416338, + 0.04491019621491432, + 0.10195999592542648, + 0.06228656694293022, + 0.21399857103824615, + 0.0426320843398571 + ], + "credit_map": [ + 0.013312694616615772, + 0.025893768295645714, + 0.018215127289295197, + 0.014760510064661503, + 0.01570909656584263, + 0.0316704697906971, + 0.02228372171521187, + 0.047264307737350464, + 0.00249200826510787, + 0.003993678372353315, + 0.0063639977015554905, + 0.00266014551743865, + 0.003908044192939997, + 0.006958333775401115, + 0.00340903177857399, + 0.002245076699182391, + 0.01196505781263113, + 0.008584748953580856, + 0.04657468944787979, + 0.006998796947300434, + 0.00583862466737628, + 0.010863028466701508, + 0.010219343937933445, + 0.04486651346087456, + 0.012910628691315651, + 0.009960904717445374, + 0.0414578877389431, + 0.012810804881155491, + 0.011361906304955482, + 0.017213953658938408, + 0.013197962194681168, + 0.03553278371691704, + 0.0029238881543278694, + 0.0014755228767171502, + 0.00144531752448529, + 0.001389196957461536, + 0.0017928717425093055, + 0.0011263269698247313, + 0.0012363281566649675, + 0.0013254671357572079, + 0.010019486770033836, + 0.005715728271752596, + 0.005287284962832928, + 0.005046519450843334, + 0.004086779430508614, + 0.004987592343240976, + 0.005418127868324518, + 0.004348673392087221, + 0.008992393501102924, + 0.02710629068315029, + 0.031804341822862625, + 0.013292090967297554, + 0.005377853289246559, + 0.004662486258894205, + 0.005634753964841366, + 0.005089789628982544, + 0.00538449315354228, + 0.014893860556185246, + 0.0201879795640707, + 0.007603020407259464, + 0.0027237003669142723, + 0.0036555961705744267, + 0.004155567381531, + 0.00368235120549798, + 0.022183476015925407, + 0.06623920798301697, + 0.07390616089105606, + 0.019821489229798317, + 0.007554288022220135, + 0.004625290632247925, + 0.012751458212733269, + 0.0069171832874417305, + 0.008444883860647678, + 0.006706658285111189, + 0.005982391536235809, + 0.008220640011131763, + 0.0031668145675212145, + 0.002797436900436878, + 0.003912318963557482, + 0.0034009392838925123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0337490439414978 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.0032168091274797916 + ], + "opposite_logprob_tokens": [ + -5.753216743469238 + ], + "logprob_reward": -0.0032168091274797916, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032168091274797916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11069401353597641, + 0.132802814245224, + 0.33353278040885925, + 0.04480068385601044, + 0.023445185273885727, + 0.02224052883684635, + 0.1757661998271942, + 0.04390750452876091, + 0.08128118515014648, + 0.03152916952967644 + ], + "credit_map": [ + 0.008825824595987797, + 0.014906680211424828, + 0.015008537098765373, + 0.007550708018243313, + 0.011704646982252598, + 0.024277646094560623, + 0.015871981158852577, + 0.012547995895147324, + 0.010480834171175957, + 0.017480120062828064, + 0.03251367434859276, + 0.006482147146016359, + 0.009339045733213425, + 0.041165921837091446, + 0.0084996921941638, + 0.006841378752142191, + 0.0256598312407732, + 0.019121669232845306, + 0.11193480342626572, + 0.015187449753284454, + 0.010927163064479828, + 0.02395988628268242, + 0.01975078694522381, + 0.10699117183685303, + 0.0036916860844939947, + 0.0033529805950820446, + 0.012531776912510395, + 0.003933837171643972, + 0.003986420575529337, + 0.004380751866847277, + 0.002361594932153821, + 0.010561637580394745, + 0.004796265624463558, + 0.0029141756240278482, + 0.0024336965288966894, + 0.002477399306371808, + 0.0030037201941013336, + 0.002014101017266512, + 0.00284747825935483, + 0.0029583489522337914, + 0.0034578817430883646, + 0.002571563236415386, + 0.002393512986600399, + 0.002619498409330845, + 0.002881864085793495, + 0.002349046990275383, + 0.003593724686652422, + 0.002373436698690057, + 0.013994433917105198, + 0.053138259798288345, + 0.056188687682151794, + 0.014411983080208302, + 0.007572025526314974, + 0.009439936839044094, + 0.0118172038346529, + 0.00920366682112217, + 0.003965491894632578, + 0.0094566335901618, + 0.010677198879420757, + 0.007633202243596315, + 0.0031478211749345064, + 0.0026165388990193605, + 0.0035195392556488514, + 0.00289108045399189, + 0.007927157916128635, + 0.025989634916186333, + 0.030741309747099876, + 0.00542293768376112, + 0.0031840356532484293, + 0.0016744546592235565, + 0.004001992288976908, + 0.0023396627511829138, + 0.004400692880153656, + 0.003107150550931692, + 0.004954850301146507, + 0.004017251543700695, + 0.003924787510186434, + 0.0031419533770531416, + 0.004875436890870333, + 0.0031070448458194733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.0032168091274797916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.026348749175667763 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348749175667763, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348749175667763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0825311541557312, + 0.12431018054485321, + 0.1398957222700119, + 0.09681888669729233, + 0.041373565793037415, + 0.04965980350971222, + 0.13604921102523804, + 0.09804674237966537, + 0.15969175100326538, + 0.07162300497293472 + ], + "credit_map": [ + 0.008718667551875114, + 0.01047319546341896, + 0.010242902673780918, + 0.00822711642831564, + 0.00798613578081131, + 0.012574910186231136, + 0.009932939894497395, + 0.014375288039445877, + 0.017970329150557518, + 0.02155822142958641, + 0.014368639327585697, + 0.01882276125252247, + 0.012876985594630241, + 0.01523621752858162, + 0.011447017081081867, + 0.012030006386339664, + 0.011973866261541843, + 0.00817977823317051, + 0.04436968266963959, + 0.006137828808277845, + 0.005599812138825655, + 0.010363839566707611, + 0.0078197680413723, + 0.04545113816857338, + 0.010031262412667274, + 0.006582720670849085, + 0.02359548956155777, + 0.00816909596323967, + 0.008433448150753975, + 0.010440437123179436, + 0.009949618950486183, + 0.019616808742284775, + 0.005607284605503082, + 0.0046449825167655945, + 0.0036494224332273006, + 0.005492864642292261, + 0.004579825326800346, + 0.0051116649992764, + 0.005127342417836189, + 0.007160178851336241, + 0.006112894508987665, + 0.005018451251089573, + 0.0055312481708824635, + 0.006887482944875956, + 0.005513111595064402, + 0.007145097944885492, + 0.008270364254713058, + 0.005181153304874897, + 0.01809663698077202, + 0.03309980779886246, + 0.03044848144054413, + 0.020144164562225342, + 0.006642632652074099, + 0.009314891882240772, + 0.008933993987739086, + 0.00936860591173172, + 0.011951686814427376, + 0.01634935289621353, + 0.02268955484032631, + 0.022464849054813385, + 0.004772118758410215, + 0.005097388289868832, + 0.010203832760453224, + 0.0045179640874266624, + 0.01471662987023592, + 0.04203629866242409, + 0.05270703136920929, + 0.01881301961839199, + 0.006907488685101271, + 0.0045740664936602116, + 0.013955987058579922, + 0.005981229245662689, + 0.009972691535949707, + 0.008913500234484673, + 0.00991380400955677, + 0.013168762437999249, + 0.006715571042150259, + 0.0058309040032327175, + 0.00941789522767067, + 0.007689879275858402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.026348749175667763 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.07000092417001724 + ], + "opposite_logprob_tokens": [ + -2.6950008869171143 + ], + "logprob_reward": -0.07000092417001724, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000092417001724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06733045727014542, + 0.17378750443458557, + 0.29475918412208557, + 0.06102495640516281, + 0.013628355227410793, + 0.027078352868556976, + 0.07380238175392151, + 0.17312097549438477, + 0.07383459061384201, + 0.04163328930735588 + ], + "credit_map": [ + 0.007366785779595375, + 0.011008079163730145, + 0.006771894171833992, + 0.007305836770683527, + 0.008323565125465393, + 0.01079409196972847, + 0.008275139145553112, + 0.0074850693345069885, + 0.012346245348453522, + 0.024243207648396492, + 0.03759583458304405, + 0.010288253426551819, + 0.011296103708446026, + 0.05791845545172691, + 0.011218016967177391, + 0.008881377056241035, + 0.022795405238866806, + 0.016563156619668007, + 0.10096516460180283, + 0.013097995892167091, + 0.009564516134560108, + 0.021138930693268776, + 0.01581423170864582, + 0.09481975436210632, + 0.005137146916240454, + 0.003670078469440341, + 0.019390078261494637, + 0.004647230263799429, + 0.003424141788855195, + 0.006288138218224049, + 0.0023082720581442118, + 0.016159869730472565, + 0.0027306058909744024, + 0.0015210313722491264, + 0.00153359433170408, + 0.0019156207563355565, + 0.0016898749163374305, + 0.0011967241298407316, + 0.0017119310796260834, + 0.0013289734488353133, + 0.00673138489946723, + 0.004349973984062672, + 0.003886840073391795, + 0.0025532261934131384, + 0.0023148232139647007, + 0.0022335874382406473, + 0.0026009553112089634, + 0.0024075605906546116, + 0.008007748983800411, + 0.01619243435561657, + 0.016213199123740196, + 0.012860204093158245, + 0.004550232086330652, + 0.003937284927815199, + 0.007447284180670977, + 0.004593989811837673, + 0.015445517376065254, + 0.028101220726966858, + 0.05295173451304436, + 0.03307413309812546, + 0.00815633125603199, + 0.01039635855704546, + 0.015560378320515156, + 0.00943529698997736, + 0.005921218078583479, + 0.019721467047929764, + 0.025473251938819885, + 0.009274262003600597, + 0.003508681198582053, + 0.0017860060324892402, + 0.005689722485840321, + 0.0024599803145974874, + 0.0067397216334939, + 0.004485113080590963, + 0.006638954859226942, + 0.008982979692518711, + 0.0032464463729411364, + 0.0025757320690900087, + 0.005829034373164177, + 0.0031353048980236053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.07000092417001724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Low-angle, static medium shots capture the moment of violent impact in breathtaking super-slow motion, rendered in high-quality cinematic detail. On a damp, bumpy cobblestone street, shrouded in mystery, a fantastical, fuel-powered tricycle collides with an unseen force. The vehicle, a masterpiece of ornate Victorian design, its frame adorned with polished brass fittings and intricate gears, gleams in the dim light. As the front of the tricycle violently collapses, the crimson velvet seats ripple with the shockwave. The true force of the impact is visualized as a nearby cast-iron gas lamp shatters, sending a cascade of glittering glass shards streaking gracefully through the air in mesmerizing slow motion. These fragments, capturing the flickering warm golden glow of the dying gas flame, seem suspended in thick diamonds as the swirling fog swirls and separates in chaos, creating a scene of intense and haunting beauty, freezing a single, devastating moment.", + "seed": 437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_3024_seed_437_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": -0.0909972014836967, + "vlm_logprob": -0.0909972014836967, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.0909972014836967, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.23620645701885223 + ], + "opposite_logprob_tokens": [ + -2.986206531524658 + ], + "logprob_reward": -0.23620645701885223, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.23620645701885223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11632487922906876, + 0.07185809314250946, + 0.21339285373687744, + 0.07807468622922897, + 0.06586574763059616, + 0.054418884217739105, + 0.17290373146533966, + 0.05984465032815933, + 0.08547817915678024, + 0.08183827251195908 + ], + "credit_map": [ + 0.01401777658611536, + 0.019511885941028595, + 0.00987318716943264, + 0.01331354584544897, + 0.014400946907699108, + 0.01697971485555172, + 0.013303471729159355, + 0.014924352057278156, + 0.010225748643279076, + 0.007879825308918953, + 0.006153261289000511, + 0.008945606648921967, + 0.008901998400688171, + 0.008778468705713749, + 0.012475459836423397, + 0.008497723378241062, + 0.013649050146341324, + 0.011506852693855762, + 0.009090360254049301, + 0.012862201780080795, + 0.007255906239151955, + 0.07000382989645004, + 0.011743214912712574, + 0.0772814229130745, + 0.008773204870522022, + 0.012833697721362114, + 0.01001131534576416, + 0.01334487646818161, + 0.009076432324945927, + 0.0070578549057245255, + 0.009550393559038639, + 0.007426914293318987, + 0.005788628943264484, + 0.009530582465231419, + 0.010307096876204014, + 0.014438841491937637, + 0.006130261812359095, + 0.004343242850154638, + 0.009041327983140945, + 0.006285764742642641, + 0.0058647263795137405, + 0.007478261832147837, + 0.010575936175882816, + 0.010513143613934517, + 0.004316544160246849, + 0.0035817958414554596, + 0.006811508443206549, + 0.005276970099657774, + 0.04836317524313927, + 0.046879272907972336, + 0.019334636628627777, + 0.026054706424474716, + 0.01053253561258316, + 0.008925061672925949, + 0.00797510240226984, + 0.004839232657104731, + 0.005793894175440073, + 0.012679388746619225, + 0.007468400057405233, + 0.011246980167925358, + 0.005101497750729322, + 0.004517042078077793, + 0.00737467547878623, + 0.005662775132805109, + 0.00849740207195282, + 0.015748219564557076, + 0.0078813461586833, + 0.012957190163433552, + 0.009256749413907528, + 0.010375255718827248, + 0.008780031464993954, + 0.011981986463069916, + 0.013836827129125595, + 0.007280639372766018, + 0.008232321590185165, + 0.010723607614636421, + 0.00814895611256361, + 0.01807035505771637, + 0.006890102755278349, + 0.008655463345348835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.23620645701885223 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.05501488596200943 + ], + "opposite_logprob_tokens": [ + -2.9300148487091064 + ], + "logprob_reward": -0.05501488596200943, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501488596200943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22894899547100067, + 0.08663561195135117, + 0.2305806279182434, + 0.08339687436819077, + 0.04964221268892288, + 0.05483555793762207, + 0.06295891851186752, + 0.06468990445137024, + 0.06823624670505524, + 0.07007504999637604 + ], + "credit_map": [ + 0.03145816922187805, + 0.04360059276223183, + 0.014553116634488106, + 0.024027500301599503, + 0.03242698684334755, + 0.035142552107572556, + 0.023755064234137535, + 0.0239850003272295, + 0.008827731013298035, + 0.015094393864274025, + 0.008833588100969791, + 0.009277968667447567, + 0.0130727244541049, + 0.01054466050118208, + 0.012402251362800598, + 0.008582284674048424, + 0.014073571190237999, + 0.012883729301393032, + 0.015886876732110977, + 0.015703639015555382, + 0.009173408150672913, + 0.06721725314855576, + 0.013945354148745537, + 0.0816967785358429, + 0.008120219223201275, + 0.01281646080315113, + 0.014434803277254105, + 0.013789848424494267, + 0.006690806243568659, + 0.007767488714307547, + 0.012460673227906227, + 0.0073165688663721085, + 0.00559291010722518, + 0.0070567200891673565, + 0.007818074896931648, + 0.00907451193779707, + 0.005744297988712788, + 0.0042789350263774395, + 0.005829332862049341, + 0.004247422330081463, + 0.006468534469604492, + 0.00747247040271759, + 0.006993341725319624, + 0.009676291607320309, + 0.005947116296738386, + 0.0040459162555634975, + 0.00888567604124546, + 0.0053462074138224125, + 0.0035415170714259148, + 0.01731928065419197, + 0.00564213702455163, + 0.022556032985448837, + 0.0038109244778752327, + 0.0026365763042122126, + 0.004762859549373388, + 0.00268958299420774, + 0.0052341194823384285, + 0.010433673858642578, + 0.010644554160535336, + 0.009810298681259155, + 0.0073262471705675125, + 0.0056709349155426025, + 0.009266605600714684, + 0.0063034649938344955, + 0.007272251881659031, + 0.010938604362308979, + 0.006499607115983963, + 0.008688844740390778, + 0.008344762027263641, + 0.00912838988006115, + 0.00857054628431797, + 0.008793232962489128, + 0.0070115430280566216, + 0.006286424119025469, + 0.011577820405364037, + 0.009028030559420586, + 0.006846246775239706, + 0.010766073130071163, + 0.006892112549394369, + 0.011666794307529926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.05501488596200943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.006768987979739904 + ], + "opposite_logprob_tokens": [ + -5.006769180297852 + ], + "logprob_reward": -0.006768987979739904, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006768987979739904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0878635048866272, + 0.06356954574584961, + 0.21387554705142975, + 0.1518491804599762, + 0.06501183658838272, + 0.0844869613647461, + 0.1735316962003708, + 0.05661512166261673, + 0.041608572006225586, + 0.061587996780872345 + ], + "credit_map": [ + 0.011220403015613556, + 0.014968061819672585, + 0.008112788200378418, + 0.009247216396033764, + 0.01162790972739458, + 0.013454322703182697, + 0.008770517073571682, + 0.010462296195328236, + 0.012076659128069878, + 0.008827384561300278, + 0.006762807723134756, + 0.005918994080275297, + 0.00773648452013731, + 0.007150585297495127, + 0.007653203792870045, + 0.007443429436534643, + 0.018050262704491615, + 0.020769353955984116, + 0.023645799607038498, + 0.015673015266656876, + 0.011472217738628387, + 0.03894393891096115, + 0.020135344937443733, + 0.06518564373254776, + 0.015404277481138706, + 0.019491389393806458, + 0.028968950733542442, + 0.024333849549293518, + 0.011082255281507969, + 0.01711788959801197, + 0.021030208095908165, + 0.014420381747186184, + 0.007471238262951374, + 0.008126456290483475, + 0.006595016922801733, + 0.01068320032209158, + 0.005825233180075884, + 0.00577561417594552, + 0.011740599758923054, + 0.008794484660029411, + 0.008364389650523663, + 0.017665347084403038, + 0.009452158585190773, + 0.015716195106506348, + 0.008931491523981094, + 0.005736206658184528, + 0.01024092361330986, + 0.008380260318517685, + 0.010649015195667744, + 0.03951647877693176, + 0.02182484045624733, + 0.06824437528848648, + 0.00883168913424015, + 0.007355534005910158, + 0.012473675422370434, + 0.004636103753000498, + 0.005389214958995581, + 0.012869011610746384, + 0.0065822601318359375, + 0.007512721698731184, + 0.006169567350298166, + 0.004902980290353298, + 0.008500723168253899, + 0.004688647110015154, + 0.003695571795105934, + 0.006214227993041277, + 0.0046612233854830265, + 0.006779983639717102, + 0.00544654531404376, + 0.006447973195463419, + 0.004144901409745216, + 0.004218148533254862, + 0.006572018377482891, + 0.012985381297767162, + 0.0047896308824419975, + 0.010146106593310833, + 0.005780158098787069, + 0.007831636816263199, + 0.006519623566418886, + 0.006963447667658329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.006768987979739904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.014201529324054718 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201529324054718, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201529324054718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03297003358602524, + 0.029185570776462555, + 0.3139162063598633, + 0.06755203008651733, + 0.056634366512298584, + 0.1334303319454193, + 0.17686085402965546, + 0.03315948322415352, + 0.04235897213220596, + 0.11393220722675323 + ], + "credit_map": [ + 0.0035733766853809357, + 0.00477673439309001, + 0.003118146676570177, + 0.004013473633676767, + 0.003154164645820856, + 0.004999874625355005, + 0.004034515470266342, + 0.005299748387187719, + 0.0033860281109809875, + 0.0046007754281163216, + 0.0026920048985630274, + 0.002938816323876381, + 0.004502298776060343, + 0.0037578793708235025, + 0.00451915105804801, + 0.002788616344332695, + 0.0181980412453413, + 0.016668567433953285, + 0.01632559299468994, + 0.017508016899228096, + 0.010054513812065125, + 0.09428936243057251, + 0.020395763218402863, + 0.12047632783651352, + 0.0070086875930428505, + 0.011318913660943508, + 0.006776721216738224, + 0.00957596767693758, + 0.006379983853548765, + 0.00753509858623147, + 0.008829427883028984, + 0.0101272352039814, + 0.0066187409684062, + 0.007462231442332268, + 0.010244907811284065, + 0.0096460971981287, + 0.006063851993530989, + 0.003829552326351404, + 0.006489807739853859, + 0.0062791770324110985, + 0.013447911478579044, + 0.03395744413137436, + 0.012576160952448845, + 0.025650568306446075, + 0.009323648177087307, + 0.00770428916439414, + 0.014591483399271965, + 0.016178825870156288, + 0.010326892137527466, + 0.04134105518460274, + 0.02174042910337448, + 0.06726470589637756, + 0.00889540184289217, + 0.008792166598141193, + 0.012518421746790409, + 0.00598179642111063, + 0.0026078077498823404, + 0.006761264055967331, + 0.003930447157472372, + 0.005051193758845329, + 0.0035580825060606003, + 0.004127634223550558, + 0.00413427734747529, + 0.0029887747950851917, + 0.004097478464245796, + 0.00703131640329957, + 0.004367725923657417, + 0.0066594150848686695, + 0.004259870853275061, + 0.00474032387137413, + 0.006948177237063646, + 0.004254663363099098, + 0.011925307102501392, + 0.0261310413479805, + 0.012497251853346825, + 0.02054331824183464, + 0.007790970616042614, + 0.011328049935400486, + 0.011452658101916313, + 0.012263604439795017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.014201529324054718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.1427941471338272 + ], + "opposite_logprob_tokens": [ + -2.017794132232666 + ], + "logprob_reward": -0.1427941471338272, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427941471338272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04230474308133125, + 0.03147497400641441, + 0.10014474391937256, + 0.02377215586602688, + 0.11106757819652557, + 0.20974163711071014, + 0.30868855118751526, + 0.0727350264787674, + 0.030344605445861816, + 0.06972604244947433 + ], + "credit_map": [ + 0.005858440883457661, + 0.007137753535062075, + 0.004311849363148212, + 0.0047820922918617725, + 0.004527396988123655, + 0.005960411857813597, + 0.005287356674671173, + 0.00443943589925766, + 0.0035762896295636892, + 0.005448062904179096, + 0.004021415486931801, + 0.003822935279458761, + 0.004465826321393251, + 0.003690281417220831, + 0.0037011548411101103, + 0.0027490057982504368, + 0.006119662895798683, + 0.006473900284618139, + 0.007716964464634657, + 0.006961541250348091, + 0.004406659863889217, + 0.02809746563434601, + 0.006721377372741699, + 0.03364716097712517, + 0.002824986819177866, + 0.0032540808897465467, + 0.003124779090285301, + 0.0033111905213445425, + 0.0022809316869825125, + 0.0025399019941687584, + 0.0031763853039592505, + 0.0032598967663943768, + 0.014922453090548515, + 0.01377857755869627, + 0.021850164979696274, + 0.016003340482711792, + 0.010931078344583511, + 0.007027782034128904, + 0.013723361305892467, + 0.012830812484025955, + 0.0257475096732378, + 0.042743101716041565, + 0.02966144122183323, + 0.03859007731080055, + 0.02071424014866352, + 0.014405055902898312, + 0.01771288365125656, + 0.020167306065559387, + 0.04910026863217354, + 0.07927912473678589, + 0.03798932582139969, + 0.08698182553052902, + 0.01484476774930954, + 0.011852803640067577, + 0.01766383834183216, + 0.0109765799716115, + 0.006858536507934332, + 0.013231111690402031, + 0.007781661581248045, + 0.015720196068286896, + 0.006803867407143116, + 0.0071319998241961, + 0.009228447452187538, + 0.0059791989624500275, + 0.002343118190765381, + 0.004217245616018772, + 0.0037895527202636003, + 0.005308778956532478, + 0.0027483904268592596, + 0.0034464094787836075, + 0.004772830288857222, + 0.0037182767409831285, + 0.006369752809405327, + 0.01549143809825182, + 0.008168297819793224, + 0.015025869011878967, + 0.0053901649080216885, + 0.006055206526070833, + 0.006728233769536018, + 0.00649707205593586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.1427941471338272 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Low-angle, static medium shots capture the moment of violent impact in breathtaking super-slow motion, rendered in high-quality cinematic detail. On a damp, bumpy cobblestone street, shrouded in mystery, a fantastical, fuel-powered tricycle collides with an unseen force. The vehicle, a masterpiece of ornate Victorian design, its frame adorned with polished brass fittings and intricate gears, gleams in the dim light. As the front of the tricycle violently collapses, the crimson velvet seats ripple with the shockwave. The true force of the impact is visualized as a nearby cast-iron gas lamp shatters, sending a cascade of glittering glass shards streaking gracefully through the air in mesmerizing slow motion. These fragments, capturing the flickering warm golden glow of the dying gas flame, seem suspended in thick diamonds as the swirling fog swirls and separates in chaos, creating a scene of intense and haunting beauty, freezing a single, devastating moment.", + "seed": 438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_3024_seed_438_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": -0.05127809862606227, + "vlm_logprob": -0.05127809862606227, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.05127809862606227, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.1750829964876175 + ], + "opposite_logprob_tokens": [ + -2.6750829219818115 + ], + "logprob_reward": -0.1750829964876175, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1750829964876175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1784217357635498, + 0.11197063326835632, + 0.09342671185731888, + 0.13013295829296112, + 0.06490611284971237, + 0.06897994130849838, + 0.12280615419149399, + 0.08408183604478836, + 0.06894691288471222, + 0.07632701843976974 + ], + "credit_map": [ + 0.02718193270266056, + 0.016084441915154457, + 0.02097904682159424, + 0.02667318284511566, + 0.018982214853167534, + 0.021345799788832664, + 0.019836345687508583, + 0.027338767424225807, + 0.011659125797450542, + 0.010447758249938488, + 0.011620310135185719, + 0.018391262739896774, + 0.008652783930301666, + 0.012076620012521744, + 0.02074800804257393, + 0.018374759703874588, + 0.015459448099136353, + 0.008650921285152435, + 0.014354849234223366, + 0.020206786692142487, + 0.006085975561290979, + 0.009014087729156017, + 0.01042497344315052, + 0.009229661896824837, + 0.014604880474507809, + 0.021042350679636, + 0.01203839760273695, + 0.013210530392825603, + 0.021395115181803703, + 0.013981934636831284, + 0.02284606173634529, + 0.011013679206371307, + 0.00841561146080494, + 0.007506385445594788, + 0.008172188885509968, + 0.00943351723253727, + 0.004680234473198652, + 0.006836658343672752, + 0.00898752361536026, + 0.010873990133404732, + 0.00732329860329628, + 0.008901011198759079, + 0.009919211268424988, + 0.007982282899320126, + 0.00886512454599142, + 0.007217825390398502, + 0.007843478582799435, + 0.010927707888185978, + 0.0212491974234581, + 0.01003279909491539, + 0.0401756688952446, + 0.009159384295344353, + 0.009279435500502586, + 0.021329732611775398, + 0.005766981281340122, + 0.005812950897961855, + 0.010775742121040821, + 0.009137291461229324, + 0.011470483615994453, + 0.010811683721840382, + 0.010031524114310741, + 0.013396777212619781, + 0.010378215461969376, + 0.008080119267106056, + 0.01052042469382286, + 0.009299075230956078, + 0.007994960062205791, + 0.00843515619635582, + 0.00885860901325941, + 0.009511264972388744, + 0.0078114718198776245, + 0.00651595089584589, + 0.017043787986040115, + 0.009887650609016418, + 0.008627654984593391, + 0.006534596439450979, + 0.009285042062401772, + 0.008229494094848633, + 0.00855357013642788, + 0.008165222592651844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.1750829964876175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.02639356628060341 + ], + "opposite_logprob_tokens": [ + -3.6513936519622803 + ], + "logprob_reward": -0.02639356628060341, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02639356628060341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12962348759174347, + 0.06802993267774582, + 0.06343008577823639, + 0.0642239972949028, + 0.22075772285461426, + 0.11033457517623901, + 0.08426552265882492, + 0.09955611824989319, + 0.05755293741822243, + 0.10222562402486801 + ], + "credit_map": [ + 0.013835594989359379, + 0.012532268650829792, + 0.0164386834949255, + 0.01793232560157776, + 0.012811729684472084, + 0.011852116324007511, + 0.02394949086010456, + 0.020271267741918564, + 0.0061163823120296, + 0.0053910864517092705, + 0.006370118353515863, + 0.010092338547110558, + 0.004377841949462891, + 0.00752252945676446, + 0.013703162781894207, + 0.014456466771662235, + 0.008985335007309914, + 0.005643488373607397, + 0.00861748680472374, + 0.012948181480169296, + 0.004409868735820055, + 0.006178590934723616, + 0.008870704099535942, + 0.007776434998959303, + 0.007630225270986557, + 0.00697834137827158, + 0.008807921782135963, + 0.007233151700347662, + 0.008445056155323982, + 0.005255707539618015, + 0.012869673781096935, + 0.007003916893154383, + 0.023580577224493027, + 0.02987823449075222, + 0.03310989961028099, + 0.040756549686193466, + 0.021245557814836502, + 0.01849433034658432, + 0.03270703926682472, + 0.020985545590519905, + 0.008531107567250729, + 0.02641364187002182, + 0.012909505516290665, + 0.014783507212996483, + 0.013216844759881496, + 0.014068251475691795, + 0.011978398077189922, + 0.00843331590294838, + 0.011480186134576797, + 0.00820021890103817, + 0.020772425457835197, + 0.007525366265326738, + 0.007105317432433367, + 0.016021884977817535, + 0.00627656327560544, + 0.0068835606798529625, + 0.010528437793254852, + 0.015268856659531593, + 0.011790744960308075, + 0.016025787219405174, + 0.016889924183487892, + 0.008704555220901966, + 0.014882765710353851, + 0.005465043243020773, + 0.009887685999274254, + 0.006622844375669956, + 0.008818858303129673, + 0.007379268296062946, + 0.009561453945934772, + 0.0059865666553378105, + 0.004550261422991753, + 0.00474599888548255, + 0.017231132835149765, + 0.017975909635424614, + 0.011449030600488186, + 0.010996348224580288, + 0.01324969157576561, + 0.011078216135501862, + 0.012099810875952244, + 0.008145484142005444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.02639356628060341 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.003221681108698249 + ], + "opposite_logprob_tokens": [ + -5.75322151184082 + ], + "logprob_reward": -0.003221681108698249, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003221681108698249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11211054027080536, + 0.222657710313797, + 0.031377341598272324, + 0.26299867033958435, + 0.10162393003702164, + 0.07837390154600143, + 0.07376547902822495, + 0.035057954490184784, + 0.04109640792012215, + 0.040938083082437515 + ], + "credit_map": [ + 0.012846235185861588, + 0.00962893944233656, + 0.01636837236583233, + 0.01636899635195732, + 0.010213085450232029, + 0.014052286744117737, + 0.01594822108745575, + 0.016684409230947495, + 0.019968627020716667, + 0.015865227207541466, + 0.023005174472928047, + 0.0269684549421072, + 0.01710415445268154, + 0.02779286354780197, + 0.045709043741226196, + 0.046244170516729355, + 0.004648680333048105, + 0.0035173019859939814, + 0.004128928761929274, + 0.0038036296609789133, + 0.002755882451310754, + 0.0031525716185569763, + 0.005382176022976637, + 0.003988172393292189, + 0.033964645117521286, + 0.029648274183273315, + 0.03154624253511429, + 0.02877684310078621, + 0.03461550176143646, + 0.02528533525764942, + 0.04812818765640259, + 0.031033655628561974, + 0.011368867941200733, + 0.015114955604076385, + 0.012134313583374023, + 0.017231326550245285, + 0.011549253016710281, + 0.008141840808093548, + 0.013463185168802738, + 0.01262018270790577, + 0.005504790227860212, + 0.012377208098769188, + 0.008375456556677818, + 0.01767759770154953, + 0.009324872866272926, + 0.0067458986304700375, + 0.010701850987970829, + 0.007666222285479307, + 0.01246577873826027, + 0.00713190296664834, + 0.021785249933600426, + 0.007022220641374588, + 0.0065515353344380856, + 0.011100227944552898, + 0.00445866584777832, + 0.003249896690249443, + 0.004518003202974796, + 0.004998458549380302, + 0.005225622095167637, + 0.004473437089473009, + 0.005724580958485603, + 0.0030858691316097975, + 0.0042639258317649364, + 0.002768058329820633, + 0.007110420614480972, + 0.0042272768914699554, + 0.0048290747217834, + 0.005411580204963684, + 0.005034530535340309, + 0.0054369778372347355, + 0.0052840812131762505, + 0.00376246590167284, + 0.00472525367513299, + 0.0041429391130805016, + 0.0048520867712795734, + 0.005832897499203682, + 0.0046411678194999695, + 0.005609840154647827, + 0.006667453330010176, + 0.004466441459953785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.003221681108698249 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.008643122389912605 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643122389912605, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643122389912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0936029776930809, + 0.18147818744182587, + 0.05237390100955963, + 0.14161363244056702, + 0.13633912801742554, + 0.08164197951555252, + 0.17478299140930176, + 0.05285616219043732, + 0.03698902949690819, + 0.04832206293940544 + ], + "credit_map": [ + 0.010124622844159603, + 0.01124117523431778, + 0.011721206828951836, + 0.010450292378664017, + 0.013919899240136147, + 0.009097286500036716, + 0.012530925683677197, + 0.014517571777105331, + 0.0222813468426466, + 0.026129208505153656, + 0.023394934833049774, + 0.025362681597471237, + 0.01753516122698784, + 0.01794610545039177, + 0.024154076352715492, + 0.024674680083990097, + 0.007416750770062208, + 0.003972787410020828, + 0.006813641171902418, + 0.005836584605276585, + 0.0061492156237363815, + 0.007202369626611471, + 0.010055331513285637, + 0.004927218426018953, + 0.0196657907217741, + 0.01960781216621399, + 0.018000369891524315, + 0.014809677377343178, + 0.01701740361750126, + 0.01529130432754755, + 0.01949971914291382, + 0.01772153563797474, + 0.013970754109323025, + 0.01624704897403717, + 0.017757974565029144, + 0.027800733223557472, + 0.014833535999059677, + 0.010358993895351887, + 0.01821175217628479, + 0.017158329486846924, + 0.009604799561202526, + 0.01312577910721302, + 0.009426800534129143, + 0.013284730724990368, + 0.010337311774492264, + 0.007142817135900259, + 0.01097147911787033, + 0.007748258765786886, + 0.0348251610994339, + 0.014141524210572243, + 0.059504248201847076, + 0.013006458058953285, + 0.01081822905689478, + 0.020810922607779503, + 0.008649188093841076, + 0.013027246110141277, + 0.006015836726874113, + 0.007514361757785082, + 0.0069687687791883945, + 0.0072331782430410385, + 0.007936486043035984, + 0.004647932946681976, + 0.007992117665708065, + 0.004547484684735537, + 0.005114820785820484, + 0.005484523717314005, + 0.003783463966101408, + 0.0050415038131177425, + 0.006312800571322441, + 0.003836902789771557, + 0.00491258455440402, + 0.0025024274364113808, + 0.005678257439285517, + 0.005755809135735035, + 0.005742189008742571, + 0.006731647532433271, + 0.005605680402368307, + 0.005807663779705763, + 0.007911105640232563, + 0.005089712794870138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.008643122389912605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.043049126863479614 + ], + "opposite_logprob_tokens": [ + -3.1680490970611572 + ], + "logprob_reward": -0.043049126863479614, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043049126863479614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05856887623667717, + 0.06099620461463928, + 0.021239561960101128, + 0.09123516082763672, + 0.24165897071361542, + 0.10513827204704285, + 0.19697590172290802, + 0.045966751873493195, + 0.052564892917871475, + 0.12565551698207855 + ], + "credit_map": [ + 0.009049084037542343, + 0.0057181091979146, + 0.008644291199743748, + 0.007369353901594877, + 0.005702979397028685, + 0.005474647041410208, + 0.008299377746880054, + 0.008311021141707897, + 0.007864654064178467, + 0.005187870003283024, + 0.006086351815611124, + 0.01029097568243742, + 0.00438929907977581, + 0.00669007608667016, + 0.010997428558766842, + 0.009489532560110092, + 0.002942947670817375, + 0.001456982223317027, + 0.002445088466629386, + 0.002551561687141657, + 0.002503958297893405, + 0.0028243064880371094, + 0.003910713363438845, + 0.0026039995718747377, + 0.013568958267569542, + 0.013067344203591347, + 0.010335614904761314, + 0.008990242145955563, + 0.01343517005443573, + 0.010415131226181984, + 0.010535885579884052, + 0.010886791162192822, + 0.0277097225189209, + 0.03242991119623184, + 0.029797831550240517, + 0.03644612431526184, + 0.02334447018802166, + 0.019091332331299782, + 0.03130810707807541, + 0.041531432420015335, + 0.01556079089641571, + 0.01026927400380373, + 0.014422602020204067, + 0.01750652678310871, + 0.011203346773982048, + 0.009475084021687508, + 0.013647899031639099, + 0.013052722439169884, + 0.029593057930469513, + 0.016656262800097466, + 0.05905425176024437, + 0.016126852482557297, + 0.015545726753771305, + 0.037413593381643295, + 0.011971899308264256, + 0.010614217258989811, + 0.008000539615750313, + 0.0044242991134524345, + 0.010773432441055775, + 0.003822578815743327, + 0.00477029150351882, + 0.005705568939447403, + 0.004022644832730293, + 0.004447387997061014, + 0.008618544787168503, + 0.005579454358667135, + 0.006539348978549242, + 0.007330299820750952, + 0.005572961643338203, + 0.007067647762596607, + 0.005246349144726992, + 0.006610272452235222, + 0.014134706929326057, + 0.020254353061318398, + 0.01340312696993351, + 0.013165130279958248, + 0.016166899353265762, + 0.014365981332957745, + 0.01673034578561783, + 0.017434947192668915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.043049126863479614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Low-angle, static medium shots capture the moment of violent impact in breathtaking super-slow motion, rendered in high-quality cinematic detail. On a damp, bumpy cobblestone street, shrouded in mystery, a fantastical, fuel-powered tricycle collides with an unseen force. The vehicle, a masterpiece of ornate Victorian design, its frame adorned with polished brass fittings and intricate gears, gleams in the dim light. As the front of the tricycle violently collapses, the crimson velvet seats ripple with the shockwave. The true force of the impact is visualized as a nearby cast-iron gas lamp shatters, sending a cascade of glittering glass shards streaking gracefully through the air in mesmerizing slow motion. These fragments, capturing the flickering warm golden glow of the dying gas flame, seem suspended in thick diamonds as the swirling fog swirls and separates in chaos, creating a scene of intense and haunting beauty, freezing a single, devastating moment.", + "seed": 439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_3/index_3024_seed_439_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": -0.07695216373540462, + "vlm_logprob": -0.07695216373540462, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.07695216373540462, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.24579909443855286 + ], + "opposite_logprob_tokens": [ + -2.2457990646362305 + ], + "logprob_reward": -0.24579909443855286, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.24579909443855286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2018691897392273, + 0.06110966578125954, + 0.05366243049502373, + 0.04587931931018829, + 0.1402624100446701, + 0.1535211056470871, + 0.16169433295726776, + 0.06686612218618393, + 0.028146425262093544, + 0.08698899298906326 + ], + "credit_map": [ + 0.018770985305309296, + 0.02920222468674183, + 0.020758099853992462, + 0.021753452718257904, + 0.023022720590233803, + 0.023966088891029358, + 0.03312378376722336, + 0.03127182275056839, + 0.005638505797833204, + 0.007530910428613424, + 0.00799044594168663, + 0.007530102040618658, + 0.009100948460400105, + 0.008141344413161278, + 0.00812948402017355, + 0.007047924678772688, + 0.004540902562439442, + 0.009007042273879051, + 0.006771592423319817, + 0.010912779718637466, + 0.004783795215189457, + 0.006970772985368967, + 0.004387608263641596, + 0.0062879337929189205, + 0.003763283137232065, + 0.005137792322784662, + 0.007784462068229914, + 0.008636954240500927, + 0.003102182177826762, + 0.005542475264519453, + 0.006647205911576748, + 0.005264966748654842, + 0.012052501551806927, + 0.01785379648208618, + 0.023539002984762192, + 0.015491284430027008, + 0.019145963713526726, + 0.022760137915611267, + 0.010540121234953403, + 0.01887960359454155, + 0.01318408828228712, + 0.01475230697542429, + 0.01986275054514408, + 0.015090684406459332, + 0.017491992563009262, + 0.026645870879292488, + 0.024273214861750603, + 0.022220207378268242, + 0.014587139710783958, + 0.03156523406505585, + 0.017388619482517242, + 0.01511539239436388, + 0.012186553329229355, + 0.028171291574835777, + 0.014352255500853062, + 0.028327856212854385, + 0.006467697210609913, + 0.010797025635838509, + 0.010592233389616013, + 0.01208915002644062, + 0.006076656747609377, + 0.007324399426579475, + 0.006768647115677595, + 0.006750311236828566, + 0.002383383922278881, + 0.003886123653501272, + 0.0037868255749344826, + 0.003428125288337469, + 0.004325647838413715, + 0.0034366652835160494, + 0.004690886475145817, + 0.0022087665274739265, + 0.011999251320958138, + 0.011185555718839169, + 0.015883995220065117, + 0.013313381932675838, + 0.008651334792375565, + 0.01105461735278368, + 0.007881589233875275, + 0.007019270211458206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.24579909443855286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.07905630767345428 + ], + "opposite_logprob_tokens": [ + -2.5790562629699707 + ], + "logprob_reward": -0.07905630767345428, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905630767345428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17597272992134094, + 0.1415998488664627, + 0.0581345371901989, + 0.06209350377321243, + 0.08970046043395996, + 0.07199235260486603, + 0.06561561673879623, + 0.14636681973934174, + 0.10584721714258194, + 0.08267685025930405 + ], + "credit_map": [ + 0.018659289926290512, + 0.02491399087011814, + 0.01590915583074093, + 0.018976569175720215, + 0.01520814374089241, + 0.02825791947543621, + 0.02525314874947071, + 0.028794512152671814, + 0.012525255791842937, + 0.0203218013048172, + 0.018351653590798378, + 0.0170577522367239, + 0.020054617896676064, + 0.018933583050966263, + 0.016192663460969925, + 0.01816253364086151, + 0.004393985960632563, + 0.0065816426649689674, + 0.008223219774663448, + 0.014134957455098629, + 0.00479026697576046, + 0.007944426499307156, + 0.004245922900736332, + 0.007820116356015205, + 0.004587994888424873, + 0.009500139392912388, + 0.010609603486955166, + 0.010495772585272789, + 0.0047393194399774075, + 0.006902347784489393, + 0.007445783820003271, + 0.00781254656612873, + 0.00665767677128315, + 0.009343309327960014, + 0.016462333500385284, + 0.011723368428647518, + 0.009863031096756458, + 0.013337073847651482, + 0.008806804195046425, + 0.01350686326622963, + 0.007798370905220509, + 0.006749757099896669, + 0.008628417737782001, + 0.007859003730118275, + 0.006939407903701067, + 0.010921294800937176, + 0.013390174135565758, + 0.0097059216350317, + 0.006967480294406414, + 0.011431287042796612, + 0.006957001052796841, + 0.00472140870988369, + 0.005694610998034477, + 0.009856779128313065, + 0.008772437460720539, + 0.01121461484581232, + 0.014445536769926548, + 0.02058127336204052, + 0.016653554514050484, + 0.023235181346535683, + 0.017049288377165794, + 0.018034925684332848, + 0.019079480320215225, + 0.017287569120526314, + 0.008485688827931881, + 0.013555662706494331, + 0.011744990013539791, + 0.011209670454263687, + 0.019809424877166748, + 0.01475511398166418, + 0.014953266829252243, + 0.011333399452269077, + 0.013454782776534557, + 0.011396926827728748, + 0.01512947864830494, + 0.012034568935632706, + 0.009007860906422138, + 0.010332879610359669, + 0.0062850527465343475, + 0.005035307724028826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.07905630767345428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.0036385541316121817 + ], + "opposite_logprob_tokens": [ + -5.628638744354248 + ], + "logprob_reward": -0.0036385541316121817, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036385541316121817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26686209440231323, + 0.11189402639865875, + 0.1061493456363678, + 0.06000876799225807, + 0.11530456691980362, + 0.06900741159915924, + 0.1273461878299713, + 0.06847693771123886, + 0.0386430099606514, + 0.036307621747255325 + ], + "credit_map": [ + 0.02510727569460869, + 0.021071478724479675, + 0.03150642290711403, + 0.030891988426446915, + 0.024841446429491043, + 0.03691043704748154, + 0.038673706352710724, + 0.05785936862230301, + 0.010149190202355385, + 0.018967023119330406, + 0.01365069579333067, + 0.016736702993512154, + 0.009538073092699051, + 0.014341320842504501, + 0.011967023834586143, + 0.016544004902243614, + 0.010666247457265854, + 0.027322236448526382, + 0.008436082862317562, + 0.01849251613020897, + 0.006584143731743097, + 0.014633873477578163, + 0.008365936577320099, + 0.011648312211036682, + 0.0040113129653036594, + 0.006602213717997074, + 0.007836542092263699, + 0.01347814965993166, + 0.005221317056566477, + 0.007873501628637314, + 0.007674466818571091, + 0.007311268709599972, + 0.010058533400297165, + 0.014567522332072258, + 0.014174791984260082, + 0.015782563015818596, + 0.014278165996074677, + 0.016183460131287575, + 0.012017998844385147, + 0.018241535872220993, + 0.0050566657446324825, + 0.008577347733080387, + 0.008730045519769192, + 0.006697650067508221, + 0.007467222400009632, + 0.015492836013436317, + 0.00816681794822216, + 0.008818828500807285, + 0.009745165705680847, + 0.014083849266171455, + 0.009888235479593277, + 0.008240650407969952, + 0.010995103977620602, + 0.031446028500795364, + 0.01820359192788601, + 0.024743571877479553, + 0.007795417681336403, + 0.010388691909611225, + 0.010256219655275345, + 0.008954271674156189, + 0.008277151733636856, + 0.008436010219156742, + 0.006899277679622173, + 0.007469904143363237, + 0.0031542954966425896, + 0.004641125909984112, + 0.004901545122265816, + 0.006947631947696209, + 0.0054511502385139465, + 0.004726327024400234, + 0.005856313277035952, + 0.0029646202456206083, + 0.004755728878080845, + 0.003746548667550087, + 0.007587900850921869, + 0.005290889646857977, + 0.0037651017773896456, + 0.005635032895952463, + 0.0033423604909330606, + 0.0021840615663677454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0036385541316121817 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11735033988952637, + 0.18047267198562622, + 0.07137569785118103, + 0.12435954064130783, + 0.19226159155368805, + 0.07572932541370392, + 0.06845664978027344, + 0.07841923087835312, + 0.0417601652443409, + 0.04981480911374092 + ], + "credit_map": [ + 0.011880130507051945, + 0.020413672551512718, + 0.011423137038946152, + 0.014517432078719139, + 0.01149863563477993, + 0.013589855283498764, + 0.01631166599690914, + 0.017715806141495705, + 0.015519533306360245, + 0.02615494094789028, + 0.021895598620176315, + 0.02163003757596016, + 0.022227294743061066, + 0.02273174375295639, + 0.025318440049886703, + 0.02499508298933506, + 0.00739307189360261, + 0.011104577220976353, + 0.007088908925652504, + 0.012506271712481976, + 0.005841677542775869, + 0.01135670579969883, + 0.007100551389157772, + 0.008983929641544819, + 0.011185109615325928, + 0.014902820810675621, + 0.0209832563996315, + 0.020658468827605247, + 0.013379162177443504, + 0.015335497446358204, + 0.011735725216567516, + 0.01617949642241001, + 0.013203282840549946, + 0.019403470680117607, + 0.01804209314286709, + 0.026024827733635902, + 0.030946874991059303, + 0.03353573754429817, + 0.022529475390911102, + 0.028575832024216652, + 0.006025876384228468, + 0.0085012037307024, + 0.011069150641560555, + 0.008704512380063534, + 0.009182014502584934, + 0.010583322495222092, + 0.011120902374386787, + 0.010542343370616436, + 0.006853255908936262, + 0.008140802383422852, + 0.011903909035027027, + 0.008590219542384148, + 0.006304887589067221, + 0.007246021647006273, + 0.00939391553401947, + 0.010023639537394047, + 0.009348119609057903, + 0.009987711906433105, + 0.015277222730219364, + 0.01146274246275425, + 0.007323655765503645, + 0.009595162235200405, + 0.007647549733519554, + 0.007777065970003605, + 0.003483566688373685, + 0.003918597474694252, + 0.0048003289848566055, + 0.006229277700185776, + 0.0050895316526293755, + 0.0048363711684942245, + 0.008871386758983135, + 0.004531103186309338, + 0.009150078520178795, + 0.005915216635912657, + 0.009662275202572346, + 0.0049458700232207775, + 0.004655288066715002, + 0.008013376034796238, + 0.004059736616909504, + 0.0034129666164517403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.007630250882357359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and is the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a damp, bumpy cobblestone street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a fantastical fuel-powered tricycle with ornate Victorian design, polished brass fittings, and intricate gears visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the front of the tricycle violently collapse while crimson velvet seats ripple with the shockwave?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a cast-iron gas lamp shown shattering into glittering glass shards that streak through the air in super-slow 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": [ + -0.04863661155104637 + ], + "opposite_logprob_tokens": [ + -3.0486366748809814 + ], + "logprob_reward": -0.04863661155104637, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863661155104637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0531541109085083, + 0.08750097453594208, + 0.05393420532345772, + 0.06515632569789886, + 0.2937833368778229, + 0.2528802156448364, + 0.06270002573728561, + 0.049870189279317856, + 0.029308592900633812, + 0.05171206593513489 + ], + "credit_map": [ + 0.0046321856789290905, + 0.0095866359770298, + 0.0036159767769277096, + 0.005055972374975681, + 0.00408850098028779, + 0.008326942101120949, + 0.007448357995599508, + 0.010399531573057175, + 0.0071767782792449, + 0.011115985922515392, + 0.012043414637446404, + 0.008825009688735008, + 0.01444285362958908, + 0.0117616206407547, + 0.010798072442412376, + 0.01133722998201847, + 0.004951351322233677, + 0.010757463984191418, + 0.0049445671029388905, + 0.007033256813883781, + 0.0042812335304915905, + 0.006451014894992113, + 0.006921015214174986, + 0.00859429594129324, + 0.005756477825343609, + 0.006327059585601091, + 0.009778431616723537, + 0.011349660344421864, + 0.005153368692845106, + 0.010003004223108292, + 0.007956895977258682, + 0.008831427432596684, + 0.022117432206869125, + 0.029087932780385017, + 0.023642849177122116, + 0.041956041008234024, + 0.03512489050626755, + 0.06621512025594711, + 0.027389047667384148, + 0.04825000837445259, + 0.016410844400525093, + 0.02549678087234497, + 0.023287078365683556, + 0.025183890014886856, + 0.030324023216962814, + 0.05680660903453827, + 0.03514011576771736, + 0.04023085534572601, + 0.00563021982088685, + 0.013645258732140064, + 0.006495497655123472, + 0.006526269018650055, + 0.004700017627328634, + 0.008237149566411972, + 0.00772809237241745, + 0.009737513959407806, + 0.0052345506846904755, + 0.006269084755331278, + 0.0073786149732768536, + 0.009999648667871952, + 0.005568291991949081, + 0.0052401283755898476, + 0.0052457707934081554, + 0.004934098105877638, + 0.002159736817702651, + 0.003746376372873783, + 0.0035491108428686857, + 0.005636984016746283, + 0.004176965449005365, + 0.003036552807316184, + 0.003934431821107864, + 0.0030684317462146282, + 0.0062326532788574696, + 0.0062097590416669846, + 0.00950036570429802, + 0.007459766231477261, + 0.006469643209129572, + 0.005881771445274353, + 0.006081354338675737, + 0.0038767477963119745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 1 + ], + "selected_reward": -0.04863661155104637 + } + ], + "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/0000131/group_4/index_tensor([3794, 3794])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_4/index_tensor([3794, 3794])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5ebe03c81663d641a1c69a3520a28d3407cfdd5e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_4/index_tensor([3794, 3794])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, circling close-up, with a steady, deliberate rhythm, arcs from left to right around the flames burning in the fireplace. The flames themselves are supported by several stout birch logs, their surfaces charred black, their fiery red cores visible through the cracks. The brilliant orange-yellow flames dance and flicker gracefully, occasionally punctuated by golden sparks that crackle and drift upward, ultimately disappearing into the dark hearth of rough, unpolished natural stone. The flames are the scene's sole primary light source, their soft glow casting warm, flickering, and dynamic shadows on the hearth's cast iron interior and the surrounding stone. The dominant tones of the image, intensely warm orange-red and gold, contrast sharply with the deep black and gray-brown background, creating an atmosphere of profound tranquility, comfort, and even hypnotic warmth.", + "seed": 14308, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_4/index_3794_seed_14308_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": -0.09880088812205941, + "vlm_logprob": -0.09880088812205941, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.09880088812205941, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2346922606229782 + ], + "opposite_logprob_tokens": [ + -1.6096922159194946 + ], + "logprob_reward": -0.2346922606229782, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2346922606229782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06522929668426514, + 0.039848923683166504, + 0.0557769276201725, + 0.04076724126935005, + 0.30043357610702515, + 0.1464642733335495, + 0.1601029336452484, + 0.03866926208138466, + 0.0740714743733406, + 0.07863608002662659 + ], + "credit_map": [ + 0.00647702207788825, + 0.010064278729259968, + 0.008907363750040531, + 0.006663417909294367, + 0.0057864380069077015, + 0.008761724457144737, + 0.007297137286514044, + 0.011271915398538113, + 0.006771113723516464, + 0.005368334706872702, + 0.005490344949066639, + 0.0022522462531924248, + 0.0035834547597914934, + 0.006143702194094658, + 0.0027870419435203075, + 0.0074526844546198845, + 0.005059010814875364, + 0.0050638890825212, + 0.004794407170265913, + 0.0048329271376132965, + 0.007388621103018522, + 0.009409555234014988, + 0.005576939322054386, + 0.01365157961845398, + 0.004185269121080637, + 0.006256605964154005, + 0.004037824459373951, + 0.0044434587471187115, + 0.003815272357314825, + 0.005253331735730171, + 0.006202222313731909, + 0.006573259364813566, + 0.038885507732629776, + 0.03064633719623089, + 0.014794057235121727, + 0.02875269204378128, + 0.03075217828154564, + 0.04868968576192856, + 0.03165769204497337, + 0.07625541090965271, + 0.018415426835417747, + 0.02009272389113903, + 0.011054126545786858, + 0.012576155364513397, + 0.024947116151452065, + 0.020740322768688202, + 0.014377810060977936, + 0.02426060289144516, + 0.01644521951675415, + 0.019124245271086693, + 0.019314128905534744, + 0.025962136685848236, + 0.018680986016988754, + 0.01859070174396038, + 0.020904356613755226, + 0.02108115702867508, + 0.004128107335418463, + 0.004877801984548569, + 0.0039509497582912445, + 0.0033060419373214245, + 0.0077560520730912685, + 0.004769428633153439, + 0.003916528075933456, + 0.005964353680610657, + 0.008511894382536411, + 0.011953580193221569, + 0.011531742289662361, + 0.008240647614002228, + 0.008181224577128887, + 0.00877055712044239, + 0.006456321105360985, + 0.010425510816276073, + 0.010807805694639683, + 0.02074207179248333, + 0.0046974122524261475, + 0.007545921485871077, + 0.012215971015393734, + 0.01073740515857935, + 0.004912716802209616, + 0.006976773031055927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.2346922606229782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636488899588585 + ], + "opposite_logprob_tokens": [ + -3.651364803314209 + ], + "logprob_reward": -0.02636488899588585, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636488899588585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06576920300722122, + 0.19294434785842896, + 0.15285535156726837, + 0.08557966351509094, + 0.28155121207237244, + 0.03577776625752449, + 0.06740832328796387, + 0.036322832107543945, + 0.05308074504137039, + 0.028710613027215004 + ], + "credit_map": [ + 0.00932762399315834, + 0.007922635413706303, + 0.006975945550948381, + 0.004807338118553162, + 0.00449325330555439, + 0.009368817321956158, + 0.00825424399226904, + 0.01461933832615614, + 0.024801719933748245, + 0.023715171962976456, + 0.018375897780060768, + 0.016878100112080574, + 0.018377281725406647, + 0.025586020201444626, + 0.02565694786608219, + 0.039553191512823105, + 0.015607425011694431, + 0.013286307454109192, + 0.011410941369831562, + 0.0119446637108922, + 0.015427229925990105, + 0.025968294590711594, + 0.015204367227852345, + 0.0440061055123806, + 0.008522003889083862, + 0.008358568884432316, + 0.010323514230549335, + 0.013158170506358147, + 0.008283022791147232, + 0.010341697372496128, + 0.014722698368132114, + 0.011869973503053188, + 0.03730999305844307, + 0.02750270999968052, + 0.013337930664420128, + 0.02595711499452591, + 0.02559868060052395, + 0.04718392714858055, + 0.026855818927288055, + 0.07780502736568451, + 0.0038001693319529295, + 0.004194842651486397, + 0.00276898848824203, + 0.004275219514966011, + 0.006817557383328676, + 0.005062930751591921, + 0.004300264175981283, + 0.004557789769023657, + 0.006502338219434023, + 0.008586714044213295, + 0.00620683329179883, + 0.01181910652667284, + 0.006589971017092466, + 0.008644725196063519, + 0.009163035079836845, + 0.00989559292793274, + 0.004944814834743738, + 0.004374636337161064, + 0.004646044224500656, + 0.0023288580123335123, + 0.006655154284089804, + 0.004987283144146204, + 0.002807965502142906, + 0.005578071344643831, + 0.005887467414140701, + 0.00996692106127739, + 0.008001367561519146, + 0.004912371747195721, + 0.006834673695266247, + 0.006425683852285147, + 0.004391246475279331, + 0.006661007646471262, + 0.003302151570096612, + 0.005166546907275915, + 0.0025520427152514458, + 0.003307328326627612, + 0.0036988176871091127, + 0.0032356150913983583, + 0.003626086749136448, + 0.003822021186351776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.02636488899588585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006369707407429814 + ], + "opposite_logprob_tokens": [ + -7.375637054443359 + ], + "logprob_reward": -0.0006369707407429814, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006369707407429814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11585531383752823, + 0.12393754720687866, + 0.0581950806081295, + 0.06954200565814972, + 0.27499476075172424, + 0.04578639194369316, + 0.054604705423116684, + 0.0557018518447876, + 0.10431256890296936, + 0.09706973284482956 + ], + "credit_map": [ + 0.012579664587974548, + 0.014485609717667103, + 0.015826838091015816, + 0.01016310229897499, + 0.011570673435926437, + 0.014893881045281887, + 0.017527667805552483, + 0.018807891756296158, + 0.014077245257794857, + 0.02469300851225853, + 0.009939202107489109, + 0.007866225205361843, + 0.013323156163096428, + 0.01647280529141426, + 0.013998273760080338, + 0.023567644879221916, + 0.005382016766816378, + 0.007239023223519325, + 0.0043435231782495975, + 0.0030721602961421013, + 0.005532658658921719, + 0.011219536885619164, + 0.003574584610760212, + 0.017831578850746155, + 0.008472435176372528, + 0.008800376206636429, + 0.006300223525613546, + 0.008170942775905132, + 0.005714643280953169, + 0.01060978602617979, + 0.007279033772647381, + 0.01419457420706749, + 0.03729063272476196, + 0.024462779983878136, + 0.013557960279285908, + 0.020049136132001877, + 0.026096826419234276, + 0.04795864224433899, + 0.023994378745555878, + 0.08158440887928009, + 0.005842266138643026, + 0.003054759232327342, + 0.003800068749114871, + 0.005307728890329599, + 0.008096512407064438, + 0.006329123396426439, + 0.004872556310147047, + 0.008483377285301685, + 0.005527297034859657, + 0.007704324088990688, + 0.006576333660632372, + 0.0053400699980556965, + 0.007277762517333031, + 0.0076411631889641285, + 0.005096662323921919, + 0.009441094473004341, + 0.0072648534551262856, + 0.008168714120984077, + 0.004690427798777819, + 0.0044725933112204075, + 0.010487931780517101, + 0.006717476062476635, + 0.005707221105694771, + 0.008192640729248524, + 0.01232501957565546, + 0.014923074282705784, + 0.018507199361920357, + 0.010962340980768204, + 0.01080521009862423, + 0.011819137260317802, + 0.00991857424378395, + 0.015052025206387043, + 0.013839512132108212, + 0.023787636309862137, + 0.0062250662595033646, + 0.006878947839140892, + 0.01538119651377201, + 0.012302440591156483, + 0.006393966730684042, + 0.012260973453521729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0006369707407429814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557342052459717 + ], + "opposite_logprob_tokens": [ + -1.6005734205245972 + ], + "logprob_reward": -0.22557342052459717, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557342052459717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12837529182434082, + 0.13867364823818207, + 0.11630759388208389, + 0.05614759773015976, + 0.1825876384973526, + 0.045019470155239105, + 0.0811697244644165, + 0.06256353110074997, + 0.08221028000116348, + 0.10694520175457001 + ], + "credit_map": [ + 0.018490714952349663, + 0.013871338218450546, + 0.0134061798453331, + 0.007877849042415619, + 0.008334034122526646, + 0.0204380564391613, + 0.01174371037632227, + 0.03421342745423317, + 0.021872611716389656, + 0.014507126063108444, + 0.013777484185993671, + 0.012985246255993843, + 0.015442096628248692, + 0.02076752856373787, + 0.015821589156985283, + 0.023499974980950356, + 0.01892976276576519, + 0.010826283134520054, + 0.006964941509068012, + 0.006887526251375675, + 0.00997986551374197, + 0.02160462737083435, + 0.01020804699510336, + 0.03090655244886875, + 0.00554070295765996, + 0.008426523767411709, + 0.005391980521380901, + 0.005881326738744974, + 0.005153126083314419, + 0.007330421358346939, + 0.008512236177921295, + 0.009911290369927883, + 0.01982002891600132, + 0.02963068149983883, + 0.011731160804629326, + 0.01706271804869175, + 0.026875171810388565, + 0.026388900354504585, + 0.020460696890950203, + 0.03061828389763832, + 0.004805428441613913, + 0.006517352536320686, + 0.0033813253976404667, + 0.004479113034904003, + 0.009777730330824852, + 0.005735883489251137, + 0.003545111045241356, + 0.006777530070394278, + 0.008506053127348423, + 0.00985739566385746, + 0.007850862108170986, + 0.014241225086152554, + 0.0076764593832194805, + 0.009918427094817162, + 0.010170742869377136, + 0.012948574498295784, + 0.00826920848339796, + 0.007775393780320883, + 0.007129893638193607, + 0.004855664912611246, + 0.012400663457810879, + 0.007438530679792166, + 0.00618639774620533, + 0.008507783524692059, + 0.009027035906910896, + 0.014333421364426613, + 0.011414043605327606, + 0.009278993122279644, + 0.00947166420519352, + 0.009551748633384705, + 0.0069769336842000484, + 0.012156447395682335, + 0.013221821747720242, + 0.02357468567788601, + 0.007663859520107508, + 0.012568389996886253, + 0.017727429047226906, + 0.013157030567526817, + 0.007824036292731762, + 0.01120795588940382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.22557342052459717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736899726092815 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736899726092815, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736899726092815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1016632616519928, + 0.053742457181215286, + 0.062963105738163, + 0.03900465741753578, + 0.2023208886384964, + 0.1591906100511551, + 0.12636807560920715, + 0.09888966381549835, + 0.087179034948349, + 0.06867822259664536 + ], + "credit_map": [ + 0.014667535200715065, + 0.0078507075086236, + 0.007798085920512676, + 0.0059835100546479225, + 0.00739149609580636, + 0.01765759475529194, + 0.012720208615064621, + 0.027594130486249924, + 0.007110019214451313, + 0.004488490056246519, + 0.009514468722045422, + 0.0033064070157706738, + 0.005768939387053251, + 0.008171292953193188, + 0.004277360625565052, + 0.011105479672551155, + 0.009083414450287819, + 0.00764510640874505, + 0.004819685127586126, + 0.005942246410995722, + 0.007458925247192383, + 0.008913629688322544, + 0.007612672634422779, + 0.011487430892884731, + 0.004562674090266228, + 0.006558786146342754, + 0.0034212444443255663, + 0.005385964177548885, + 0.003469319548457861, + 0.005187553353607655, + 0.004983468446880579, + 0.005435644648969173, + 0.023099452257156372, + 0.0243795458227396, + 0.011114435270428658, + 0.023481227457523346, + 0.02079503983259201, + 0.030816907063126564, + 0.021132512018084526, + 0.04750176519155502, + 0.025496695190668106, + 0.01121953222900629, + 0.015363938175141811, + 0.01580948941409588, + 0.019489992409944534, + 0.021690839901566505, + 0.015258931554853916, + 0.03486119955778122, + 0.019516749307513237, + 0.010748250409960747, + 0.011424987576901913, + 0.016187606379389763, + 0.010794500820338726, + 0.017834048718214035, + 0.011836892925202847, + 0.028025047853589058, + 0.01138948742300272, + 0.014578896574676037, + 0.007905296050012112, + 0.00787033885717392, + 0.020693743601441383, + 0.01138091180473566, + 0.01047852635383606, + 0.014592460356652737, + 0.010167174972593784, + 0.016302281990647316, + 0.00912722758948803, + 0.012058118358254433, + 0.014374279417097569, + 0.009150469675660133, + 0.006966383196413517, + 0.00903309602290392, + 0.00913033913820982, + 0.012792646884918213, + 0.006376694887876511, + 0.007946129888296127, + 0.009310104884207249, + 0.008582848124206066, + 0.005735669285058975, + 0.008803785778582096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.006736899726092815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, circling close-up, with a steady, deliberate rhythm, arcs from left to right around the flames burning in the fireplace. The flames themselves are supported by several stout birch logs, their surfaces charred black, their fiery red cores visible through the cracks. The brilliant orange-yellow flames dance and flicker gracefully, occasionally punctuated by golden sparks that crackle and drift upward, ultimately disappearing into the dark hearth of rough, unpolished natural stone. The flames are the scene's sole primary light source, their soft glow casting warm, flickering, and dynamic shadows on the hearth's cast iron interior and the surrounding stone. The dominant tones of the image, intensely warm orange-red and gold, contrast sharply with the deep black and gray-brown background, creating an atmosphere of profound tranquility, comfort, and even hypnotic warmth.", + "seed": 14309, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_4/index_3794_seed_14309_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": -0.12501969498116522, + "vlm_logprob": -0.12501969498116522, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.12501969498116522, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35724207758903503 + ], + "opposite_logprob_tokens": [ + -1.2322421073913574 + ], + "logprob_reward": -0.35724207758903503, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35724207758903503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02069094404578209, + 0.016853757202625275, + 0.0167207270860672, + 0.005065304227173328, + 0.014609545469284058, + 0.04220200330018997, + 0.020562944933772087, + 0.008765080012381077, + 0.12797708809375763, + 0.7265526056289673 + ], + "credit_map": [ + 0.003192905569449067, + 0.002614153316244483, + 0.0034259678795933723, + 0.0018120212480425835, + 0.0025991180445998907, + 0.003225761465728283, + 0.0017461454262956977, + 0.002074869815260172, + 0.0027980131562799215, + 0.002691591391339898, + 0.0016940523637458682, + 0.0014693487901240587, + 0.0030994764529168606, + 0.0020300785545259714, + 0.0013553273165598512, + 0.001715867780148983, + 0.0028540068306028843, + 0.002663564635440707, + 0.002858905354514718, + 0.001137826475314796, + 0.0009964548517018557, + 0.0021767723374068737, + 0.0010418856982141733, + 0.002991309855133295, + 0.0007738191052339971, + 0.0006708819419145584, + 0.0006721437093801796, + 0.0006238662172108889, + 0.00046844236203469336, + 0.0007293705129995942, + 0.0005084172589704394, + 0.0006183625664561987, + 0.002041162922978401, + 0.001925218035466969, + 0.00151873379945755, + 0.0018567295046523213, + 0.0030913434457033873, + 0.0019435882568359375, + 0.0008890589815564454, + 0.001343709067441523, + 0.00653162831440568, + 0.005250978283584118, + 0.003807411063462496, + 0.00413148757070303, + 0.006940067745745182, + 0.005784120876342058, + 0.005098151508718729, + 0.00465815793722868, + 0.0024672069121152163, + 0.003107865108177066, + 0.0032966318540275097, + 0.002311233663931489, + 0.002145631704479456, + 0.002771839965134859, + 0.002355610951781273, + 0.002106922911480069, + 0.0016458957688882947, + 0.0014460255624726415, + 0.0009427031618542969, + 0.0007215198129415512, + 0.0009236223995685577, + 0.0015716101042926311, + 0.0008669724338687956, + 0.0006467310595326126, + 0.02032884769141674, + 0.020350851118564606, + 0.014416505582630634, + 0.013652350753545761, + 0.016542349010705948, + 0.017419960349798203, + 0.011544119566679, + 0.013722110539674759, + 0.2942235767841339, + 0.29941046237945557, + 0.028574679046869278, + 0.016669368371367455, + 0.028036803007125854, + 0.025220297276973724, + 0.010923895053565502, + 0.023493478074669838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.35724207758903503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029826786369085312 + ], + "opposite_logprob_tokens": [ + -3.5298268795013428 + ], + "logprob_reward": -0.029826786369085312, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029826786369085312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02979528345167637, + 0.028417447581887245, + 0.049167077988386154, + 0.0413360670208931, + 0.07048032432794571, + 0.017065994441509247, + 0.017616095021367073, + 0.029848581179976463, + 0.03976112976670265, + 0.6765120625495911 + ], + "credit_map": [ + 0.004863548558205366, + 0.0032308984082192183, + 0.004564194008708, + 0.0032713741529732943, + 0.0035191711504012346, + 0.004389611072838306, + 0.002957869553938508, + 0.0029986149165779352, + 0.004255819600075483, + 0.0026882297825068235, + 0.004540395922958851, + 0.002935250988230109, + 0.0035639056004583836, + 0.0038191587664186954, + 0.002506053540855646, + 0.00410863570868969, + 0.008391489274799824, + 0.005769037641584873, + 0.0048108347691595554, + 0.0041376142762601376, + 0.005683414172381163, + 0.007332867477089167, + 0.003986905794590712, + 0.009054915979504585, + 0.007079523988068104, + 0.004607062321156263, + 0.005405878182500601, + 0.004459455143660307, + 0.00488663325086236, + 0.006833048537373543, + 0.0022655241191387177, + 0.0057989428751170635, + 0.009585599415004253, + 0.011357712559401989, + 0.005247130990028381, + 0.004940189886838198, + 0.017906323075294495, + 0.008319028653204441, + 0.0037659720983356237, + 0.009358362294733524, + 0.002594113815575838, + 0.002469339407980442, + 0.0016191183822229505, + 0.0015008861664682627, + 0.002805492142215371, + 0.0023489580489695072, + 0.0011307729873806238, + 0.0025973129086196423, + 0.00200759363360703, + 0.0030197715386748314, + 0.0018004800658673048, + 0.00195531384088099, + 0.0026325301732867956, + 0.002535480074584484, + 0.0020123173017054796, + 0.001652608043514192, + 0.004981588572263718, + 0.003767209593206644, + 0.003934864420443773, + 0.0021685692481696606, + 0.0036481160204857588, + 0.00600042100995779, + 0.0034473484847694635, + 0.0019004610367119312, + 0.006485764868557453, + 0.006830550264567137, + 0.00560472859069705, + 0.005547774489969015, + 0.006229368969798088, + 0.004850005730986595, + 0.0020720106549561024, + 0.0021409248001873493, + 0.27532142400741577, + 0.2810090184211731, + 0.026439256966114044, + 0.01476320717483759, + 0.024766966700553894, + 0.023432394489645958, + 0.008552324026823044, + 0.022227464243769646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.029826786369085312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013399679446592927 + ], + "opposite_logprob_tokens": [ + -6.626339912414551 + ], + "logprob_reward": -0.0013399679446592927, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013399679446592927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054080020636320114, + 0.02521691843867302, + 0.03973160311579704, + 0.040782395750284195, + 0.12478746473789215, + 0.028441688045859337, + 0.026167046278715134, + 0.018704060465097427, + 0.061924297362565994, + 0.5801645517349243 + ], + "credit_map": [ + 0.009003509767353535, + 0.006697004195302725, + 0.006611653603613377, + 0.002968009328469634, + 0.009766947478055954, + 0.008070559240877628, + 0.0049562836065888405, + 0.0060060517862439156, + 0.0031733540818095207, + 0.0026602428406476974, + 0.0031433559488505125, + 0.00229256646707654, + 0.005282025318592787, + 0.002809508703649044, + 0.0019369048532098532, + 0.003918960690498352, + 0.0063566118478775024, + 0.005533822812139988, + 0.004654981195926666, + 0.004054634366184473, + 0.006888061761856079, + 0.0054995594546198845, + 0.003019359428435564, + 0.003724571317434311, + 0.007461392320692539, + 0.004757028538733721, + 0.004014528822153807, + 0.003417489817366004, + 0.005436428822577, + 0.008611949160695076, + 0.002226337557658553, + 0.004857239779084921, + 0.017634274438023567, + 0.02220103144645691, + 0.009191302582621574, + 0.011072336696088314, + 0.03147546574473381, + 0.014856453984975815, + 0.005602237768471241, + 0.012754369527101517, + 0.004384136293083429, + 0.004691406153142452, + 0.002407789696007967, + 0.003021552227437496, + 0.005667046643793583, + 0.0032126354053616524, + 0.00295881531201303, + 0.0020983063150197268, + 0.0032933936454355717, + 0.004136158153414726, + 0.0037726520095020533, + 0.002986821811646223, + 0.002676169853657484, + 0.003979031927883625, + 0.003181030508130789, + 0.002141787437722087, + 0.00189789233263582, + 0.0032254105899482965, + 0.0026091879699379206, + 0.0015881845029070973, + 0.0027455780655145645, + 0.0023221054580062628, + 0.00261567160487175, + 0.0017000300576910377, + 0.009178966283798218, + 0.009524627588689327, + 0.008356601931154728, + 0.008631807751953602, + 0.006776713766157627, + 0.009441787376999855, + 0.00515605928376317, + 0.004857735708355904, + 0.23523516952991486, + 0.24163801968097687, + 0.02381731942296028, + 0.012354877777397633, + 0.02122380957007408, + 0.020099621266126633, + 0.007099722512066364, + 0.01869601011276245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.0013399679446592927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256065309047699 + ], + "opposite_logprob_tokens": [ + -1.6006065607070923 + ], + "logprob_reward": -0.2256065309047699, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256065309047699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03117132931947708, + 0.012238975614309311, + 0.017049388960003853, + 0.010873092338442802, + 0.05194799602031708, + 0.03230425715446472, + 0.012117300182580948, + 0.012300349771976471, + 0.11789090186357498, + 0.7021064162254333 + ], + "credit_map": [ + 0.00542123569175601, + 0.003989670891314745, + 0.004705285653471947, + 0.002953213406726718, + 0.003647484118118882, + 0.005245072767138481, + 0.002094479976221919, + 0.0031148847192525864, + 0.002246302552521229, + 0.0016299603739753366, + 0.0017329263500869274, + 0.0013713225489482284, + 0.001399163156747818, + 0.0016132219461724162, + 0.000967052998021245, + 0.0012790237087756395, + 0.002820225665345788, + 0.002011595293879509, + 0.0018546051578596234, + 0.0014892477774992585, + 0.0016634942730888724, + 0.002815787447616458, + 0.0016353472601622343, + 0.0027590857353061438, + 0.002280336571857333, + 0.0010071692522615194, + 0.0009258027421310544, + 0.0008759595220908523, + 0.00111028749961406, + 0.0024865835439413786, + 0.0006489056977443397, + 0.001538046170026064, + 0.0093980198726058, + 0.006208211649209261, + 0.005671977996826172, + 0.005368278827518225, + 0.006169332191348076, + 0.008391954004764557, + 0.005218994803726673, + 0.005521224811673164, + 0.004638634622097015, + 0.004630191717296839, + 0.0026191005017608404, + 0.002866655820980668, + 0.005158642306923866, + 0.004408164881169796, + 0.0021228278055787086, + 0.005860039498656988, + 0.001754491706378758, + 0.002091263420879841, + 0.0017090407200157642, + 0.0012107847724109888, + 0.0014475253410637379, + 0.0017009916482493281, + 0.001290876534767449, + 0.0009123256313614547, + 0.0019312802469357848, + 0.0015925593907013535, + 0.0011703449999913573, + 0.001150145079009235, + 0.0014222088502719998, + 0.0023823208175599575, + 0.001732938690111041, + 0.0009185513481497765, + 0.019600288942456245, + 0.023644890636205673, + 0.012193920090794563, + 0.014026915654540062, + 0.01568644866347313, + 0.013110603205859661, + 0.009648741222918034, + 0.009979087859392166, + 0.28424715995788574, + 0.28999561071395874, + 0.028249576687812805, + 0.01552171166986227, + 0.026589449495077133, + 0.02459653466939926, + 0.009954522363841534, + 0.022951846942305565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.2256065309047699 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083112098276615 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083112098276615, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083112098276615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05924668163061142, + 0.0500982329249382, + 0.03635168448090553, + 0.02419082261621952, + 0.08116456121206284, + 0.07909119129180908, + 0.043915681540966034, + 0.03770621120929718, + 0.2982746958732605, + 0.28996017575263977 + ], + "credit_map": [ + 0.01033144909888506, + 0.008131220005452633, + 0.00826802384108305, + 0.004645375069230795, + 0.005473366938531399, + 0.009306661784648895, + 0.006116101983934641, + 0.006974493153393269, + 0.007632820401340723, + 0.007134466432034969, + 0.006625122856348753, + 0.004498936701565981, + 0.00784370768815279, + 0.0056531718000769615, + 0.00531057408079505, + 0.005399436689913273, + 0.006391954608261585, + 0.004637524019926786, + 0.004367000889033079, + 0.004424611106514931, + 0.002462083240970969, + 0.005516883917152882, + 0.002280929358676076, + 0.006270701996982098, + 0.0038749976083636284, + 0.002877707127481699, + 0.002145577920600772, + 0.0016818856820464134, + 0.002721667056903243, + 0.006394607480615377, + 0.001082419534213841, + 0.0034119640477001667, + 0.012360778637230396, + 0.013597028329968452, + 0.006763692945241928, + 0.007264191284775734, + 0.0198026392608881, + 0.0106688616797328, + 0.003275937167927623, + 0.00743144191801548, + 0.013033881783485413, + 0.01461375318467617, + 0.006495155859738588, + 0.007016308140009642, + 0.01209342759102583, + 0.008775975555181503, + 0.008067634887993336, + 0.00899505615234375, + 0.005489007104188204, + 0.0068403384648263454, + 0.007727060467004776, + 0.0050710514187812805, + 0.004384947940707207, + 0.006733364425599575, + 0.003650518599897623, + 0.0040193963795900345, + 0.005313826259225607, + 0.0056751798838377, + 0.00766692915931344, + 0.0024607789237052202, + 0.004704748280346394, + 0.006584604270756245, + 0.00279610906727612, + 0.002504040952771902, + 0.04788082465529442, + 0.052367351949214935, + 0.03299446403980255, + 0.030609436333179474, + 0.04028284177184105, + 0.03912778198719025, + 0.021136535331606865, + 0.03387546166777611, + 0.11696095764636993, + 0.12039457261562347, + 0.012464587576687336, + 0.006416634656488895, + 0.010516788810491562, + 0.009860683232545853, + 0.003997223451733589, + 0.00934876874089241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.011083112098276615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, circling close-up, with a steady, deliberate rhythm, arcs from left to right around the flames burning in the fireplace. The flames themselves are supported by several stout birch logs, their surfaces charred black, their fiery red cores visible through the cracks. The brilliant orange-yellow flames dance and flicker gracefully, occasionally punctuated by golden sparks that crackle and drift upward, ultimately disappearing into the dark hearth of rough, unpolished natural stone. The flames are the scene's sole primary light source, their soft glow casting warm, flickering, and dynamic shadows on the hearth's cast iron interior and the surrounding stone. The dominant tones of the image, intensely warm orange-red and gold, contrast sharply with the deep black and gray-brown background, creating an atmosphere of profound tranquility, comfort, and even hypnotic warmth.", + "seed": 14310, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_4/index_3794_seed_14310_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": -0.17474296193104238, + "vlm_logprob": -0.17474296193104238, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.17474296193104238, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.530259370803833 + ], + "opposite_logprob_tokens": [ + -0.905259370803833 + ], + "logprob_reward": -0.530259370803833, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.530259370803833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12750986218452454, + 0.08354918658733368, + 0.23002074658870697, + 0.06404352188110352, + 0.05118225887417793, + 0.12512648105621338, + 0.03774069994688034, + 0.07706303149461746, + 0.02717825584113598, + 0.17658594250679016 + ], + "credit_map": [ + 0.014475110918283463, + 0.016988631337881088, + 0.012289147824048996, + 0.011793477460741997, + 0.025267435237765312, + 0.015315735712647438, + 0.009506801143288612, + 0.02187352441251278, + 0.008061788976192474, + 0.011503524146974087, + 0.009790531359612942, + 0.00774036580696702, + 0.017010023817420006, + 0.009624874219298363, + 0.00726567767560482, + 0.012552402913570404, + 0.027806542813777924, + 0.024924054741859436, + 0.024415919557213783, + 0.031142398715019226, + 0.03179169073700905, + 0.033136483281850815, + 0.019035691395401955, + 0.037767961621284485, + 0.008313159458339214, + 0.007466966286301613, + 0.008242209441959858, + 0.008427572436630726, + 0.007707580458372831, + 0.008073982782661915, + 0.005352364853024483, + 0.010459686629474163, + 0.005240663420408964, + 0.004577849060297012, + 0.00635768286883831, + 0.007012538146227598, + 0.006709029898047447, + 0.006742535624653101, + 0.004527901764959097, + 0.010014054365456104, + 0.012905744835734367, + 0.014889714308083057, + 0.01916549727320671, + 0.013875853270292282, + 0.019238142296671867, + 0.01683017797768116, + 0.012780648656189442, + 0.015440701507031918, + 0.0045072282664477825, + 0.003967082127928734, + 0.006128504406660795, + 0.003420326393097639, + 0.004340270068496466, + 0.0050415461882948875, + 0.0024619740433990955, + 0.007873768918216228, + 0.011652888730168343, + 0.010645049624145031, + 0.007836335338652134, + 0.008918236009776592, + 0.012006665579974651, + 0.011239366605877876, + 0.007071962114423513, + 0.007692527025938034, + 0.0040973154827952385, + 0.004217646084725857, + 0.0037283594720065594, + 0.00254560774192214, + 0.004922631196677685, + 0.0035606950987130404, + 0.0016958892811089754, + 0.0024101093877106905, + 0.03403277322649956, + 0.023849794641137123, + 0.013600189238786697, + 0.018832705914974213, + 0.016504375264048576, + 0.04002287611365318, + 0.013035613112151623, + 0.016707632690668106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.530259370803833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330484427511692 + ], + "opposite_logprob_tokens": [ + -3.7733049392700195 + ], + "logprob_reward": -0.02330484427511692, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330484427511692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1741274893283844, + 0.12743596732616425, + 0.09257101267576218, + 0.060869477689266205, + 0.06866677850484848, + 0.053507354110479355, + 0.06517338752746582, + 0.1514158397912979, + 0.032651498913764954, + 0.17358122766017914 + ], + "credit_map": [ + 0.016499634832143784, + 0.015537929721176624, + 0.023819290101528168, + 0.01847895048558712, + 0.03490553796291351, + 0.02099383808672428, + 0.02009694278240204, + 0.02379537560045719, + 0.014467221684753895, + 0.019013162702322006, + 0.015397143550217152, + 0.01533377356827259, + 0.017019959166646004, + 0.013761639595031738, + 0.011677750386297703, + 0.02076532132923603, + 0.008501884527504444, + 0.014641350135207176, + 0.011006356216967106, + 0.011469589546322823, + 0.010401097126305103, + 0.013348099775612354, + 0.0060492088086903095, + 0.017153425142169, + 0.008414560928940773, + 0.007032095454633236, + 0.006742412224411964, + 0.006628087256103754, + 0.009095067158341408, + 0.007199730258435011, + 0.0056649064645171165, + 0.010092618875205517, + 0.005581124220043421, + 0.006853989791125059, + 0.007184682879596949, + 0.007773577235639095, + 0.010752321220934391, + 0.010702421888709068, + 0.006110919173806906, + 0.013707743026316166, + 0.006928534246981144, + 0.005494555924087763, + 0.008630162104964256, + 0.0078659662976861, + 0.00730764027684927, + 0.007552426308393478, + 0.004840900655835867, + 0.004887174349278212, + 0.007512619718909264, + 0.009903330355882645, + 0.007874513044953346, + 0.005049338098615408, + 0.00850415788590908, + 0.008966857567429543, + 0.003474449273198843, + 0.013888117857277393, + 0.019006915390491486, + 0.019719740375876427, + 0.018739573657512665, + 0.02070961333811283, + 0.023101504892110825, + 0.02245914191007614, + 0.012448098510503769, + 0.015231231227517128, + 0.005190995056182146, + 0.003998520318418741, + 0.0033410091418772936, + 0.0023055560886859894, + 0.009128314442932606, + 0.003810912137851119, + 0.0015495691914111376, + 0.0033266255632042885, + 0.042250245809555054, + 0.012212215922772884, + 0.012831581756472588, + 0.01611907407641411, + 0.013103130273520947, + 0.049851782619953156, + 0.009295015595853329, + 0.0179181769490242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02330484427511692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007190502947196364 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007190502947196364, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007190502947196364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18528831005096436, + 0.08358985930681229, + 0.09084577113389969, + 0.0769045278429985, + 0.07694358378648758, + 0.0804377943277359, + 0.08755888044834137, + 0.06037477031350136, + 0.055064231157302856, + 0.20299232006072998 + ], + "credit_map": [ + 0.0193488709628582, + 0.020274225622415543, + 0.025436555966734886, + 0.028648197650909424, + 0.03056635893881321, + 0.021149273961782455, + 0.01642800308763981, + 0.023436805233359337, + 0.008309661410748959, + 0.01320952083915472, + 0.010195831768214703, + 0.00818054098635912, + 0.014904104173183441, + 0.009525703266263008, + 0.008169283159077168, + 0.011095210909843445, + 0.014258168637752533, + 0.01220613345503807, + 0.00917868036776781, + 0.01146959513425827, + 0.011240546591579914, + 0.015080389566719532, + 0.008026530966162682, + 0.009385719895362854, + 0.00898240227252245, + 0.013458720408380032, + 0.007911033928394318, + 0.008296404033899307, + 0.008277473971247673, + 0.010157015174627304, + 0.006421202793717384, + 0.013400266878306866, + 0.007110878359526396, + 0.005645114928483963, + 0.009500490501523018, + 0.007228579372167587, + 0.011552933603525162, + 0.011925658211112022, + 0.005499496124684811, + 0.01848042570054531, + 0.009021339006721973, + 0.011508884839713573, + 0.009897307492792606, + 0.006621202453970909, + 0.015362877398729324, + 0.011266392655670643, + 0.0069288043305277824, + 0.009830980561673641, + 0.009251285344362259, + 0.01588369347155094, + 0.012762455269694328, + 0.00848342478275299, + 0.010963288135826588, + 0.011626756750047207, + 0.005301699973642826, + 0.013286269269883633, + 0.008067824877798557, + 0.009446242824196815, + 0.005376478191465139, + 0.005463338457047939, + 0.009514656849205494, + 0.007656513713300228, + 0.005560620687901974, + 0.0092890914529562, + 0.008396929129958153, + 0.011122623458504677, + 0.006834887899458408, + 0.005458406172692776, + 0.00713002635166049, + 0.007286247331649065, + 0.003161264816299081, + 0.005673842038959265, + 0.04183493182063103, + 0.02800198271870613, + 0.016607262194156647, + 0.013950224965810776, + 0.022566722705960274, + 0.051905907690525055, + 0.009936509653925896, + 0.018188757821917534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0007190502947196364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134823739528656 + ], + "opposite_logprob_tokens": [ + -1.313482403755188 + ], + "logprob_reward": -0.3134823739528656, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134823739528656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17917078733444214, + 0.0953289344906807, + 0.13194292783737183, + 0.04179362952709198, + 0.05205116420984268, + 0.12135382741689682, + 0.13052307069301605, + 0.06685877591371536, + 0.0321698822081089, + 0.14880695939064026 + ], + "credit_map": [ + 0.02007186785340309, + 0.014017404057085514, + 0.0260353684425354, + 0.019469937309622765, + 0.022293737158179283, + 0.022847911342978477, + 0.022057516500353813, + 0.03237704932689667, + 0.00984447356313467, + 0.013798736035823822, + 0.009371032007038593, + 0.011415466666221619, + 0.019414672628045082, + 0.009668679907917976, + 0.010301707312464714, + 0.011514165438711643, + 0.0134528037160635, + 0.01590132527053356, + 0.013091580010950565, + 0.013793189078569412, + 0.021738069131970406, + 0.017672404646873474, + 0.011107584461569786, + 0.025185978040099144, + 0.005500073544681072, + 0.005852826870977879, + 0.003850142937153578, + 0.005414983723312616, + 0.004022322595119476, + 0.005809621885418892, + 0.004651191178709269, + 0.006692470517009497, + 0.005144279915839434, + 0.005216472782194614, + 0.004922917112708092, + 0.0061631337739527225, + 0.0067173754796385765, + 0.0077280811965465546, + 0.005384068004786968, + 0.010774830356240273, + 0.020662883296608925, + 0.0146248210221529, + 0.016917407512664795, + 0.010688669979572296, + 0.019818250089883804, + 0.01749301329255104, + 0.009422393515706062, + 0.011726394295692444, + 0.01316494308412075, + 0.024299602955579758, + 0.016020838171243668, + 0.009779667481780052, + 0.019460070878267288, + 0.017099276185035706, + 0.0066678933799266815, + 0.024030776694417, + 0.009633121080696583, + 0.010074148885905743, + 0.00784184131771326, + 0.007824164815247059, + 0.009008986875414848, + 0.009430556558072567, + 0.0060528176836669445, + 0.006993136368691921, + 0.00433382811024785, + 0.004203394055366516, + 0.0037396142724901438, + 0.0025451979599893093, + 0.008743497543036938, + 0.0035061475355178118, + 0.0016220392426475883, + 0.00347616383805871, + 0.03797146677970886, + 0.00945979543030262, + 0.009925860911607742, + 0.013284486718475819, + 0.009732116013765335, + 0.04573303461074829, + 0.007217039354145527, + 0.015483160503208637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.3134823739528656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949170328676701 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949170328676701, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949170328676701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15918508172035217, + 0.09445967525243759, + 0.12505151331424713, + 0.10109222680330276, + 0.11877553910017014, + 0.07661077380180359, + 0.08927768468856812, + 0.07977306842803955, + 0.05459970608353615, + 0.1011747345328331 + ], + "credit_map": [ + 0.01838146522641182, + 0.014872541651129723, + 0.022567106410861015, + 0.013533489778637886, + 0.027493558824062347, + 0.020018620416522026, + 0.014835545793175697, + 0.027482736855745316, + 0.009085331112146378, + 0.012251125648617744, + 0.014747132547199726, + 0.010036587715148926, + 0.014193245209753513, + 0.010155027732253075, + 0.00777515908703208, + 0.016216063871979713, + 0.017080256715416908, + 0.01809442602097988, + 0.01104168500751257, + 0.008839774876832962, + 0.015943508595228195, + 0.022907696664333344, + 0.008041578345000744, + 0.023102590814232826, + 0.014656621031463146, + 0.00899018719792366, + 0.014059578068554401, + 0.012551937252283096, + 0.008323903195559978, + 0.01584080420434475, + 0.007041643373668194, + 0.019627539440989494, + 0.012529018335044384, + 0.007384713739156723, + 0.01230988185852766, + 0.008604306727647781, + 0.01289619691669941, + 0.02099229022860527, + 0.006371747702360153, + 0.03768737241625786, + 0.010760648176074028, + 0.010611261241137981, + 0.009985530748963356, + 0.00565456273034215, + 0.012760689482092857, + 0.010861341841518879, + 0.006273915525525808, + 0.009702817536890507, + 0.008828893303871155, + 0.016447613015770912, + 0.010955377481877804, + 0.009081955067813396, + 0.012808098457753658, + 0.011391036212444305, + 0.0066526527516543865, + 0.013112049549818039, + 0.009982511401176453, + 0.00931292399764061, + 0.010037745349109173, + 0.00908118486404419, + 0.011481412686407566, + 0.011332757771015167, + 0.00810169242322445, + 0.010442836210131645, + 0.008985715918242931, + 0.0057446653954684734, + 0.007152318023145199, + 0.0051322453655302525, + 0.010036756284534931, + 0.00628203060477972, + 0.003291231580078602, + 0.007974740117788315, + 0.015372896566987038, + 0.012365278787910938, + 0.0151102589443326, + 0.011031475849449635, + 0.007519109640270472, + 0.017625300213694572, + 0.010836353525519371, + 0.011314056813716888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.005949170328676701 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, circling close-up, with a steady, deliberate rhythm, arcs from left to right around the flames burning in the fireplace. The flames themselves are supported by several stout birch logs, their surfaces charred black, their fiery red cores visible through the cracks. The brilliant orange-yellow flames dance and flicker gracefully, occasionally punctuated by golden sparks that crackle and drift upward, ultimately disappearing into the dark hearth of rough, unpolished natural stone. The flames are the scene's sole primary light source, their soft glow casting warm, flickering, and dynamic shadows on the hearth's cast iron interior and the surrounding stone. The dominant tones of the image, intensely warm orange-red and gold, contrast sharply with the deep black and gray-brown background, creating an atmosphere of profound tranquility, comfort, and even hypnotic warmth.", + "seed": 14311, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_4/index_3794_seed_14311_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": -0.06556566812796519, + "vlm_logprob": -0.06556566812796519, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.06556566812796519, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13644374907016754 + ], + "opposite_logprob_tokens": [ + -2.136443853378296 + ], + "logprob_reward": -0.13644374907016754, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13644374907016754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5627981424331665, + 0.09776701033115387, + 0.1118701696395874, + 0.018064014613628387, + 0.04439888149499893, + 0.03606109693646431, + 0.06239082291722298, + 0.015277871862053871, + 0.018045004457235336, + 0.03332693874835968 + ], + "credit_map": [ + 0.021120047196745872, + 0.007469967007637024, + 0.05755443498492241, + 0.009394703432917595, + 0.035237424075603485, + 0.2505471110343933, + 0.011010104790329933, + 0.1704643964767456, + 0.008723322302103043, + 0.014238539151847363, + 0.010655087418854237, + 0.006115001160651445, + 0.011853407137095928, + 0.020265739411115646, + 0.007820972241461277, + 0.018094945698976517, + 0.014640702866017818, + 0.014471346512436867, + 0.014900369569659233, + 0.009501025080680847, + 0.010920330882072449, + 0.013031961396336555, + 0.01780058443546295, + 0.016603849828243256, + 0.0025841491296887398, + 0.0026139281690120697, + 0.0027079686988145113, + 0.001973214326426387, + 0.0015526411589235067, + 0.002484631957486272, + 0.0019134371541440487, + 0.0022340440191328526, + 0.00660189613699913, + 0.005845389794558287, + 0.005350749474018812, + 0.0047876592725515366, + 0.005829247180372477, + 0.0056150187738239765, + 0.003952056635171175, + 0.0064168632961809635, + 0.004079694859683514, + 0.004287830088287592, + 0.0051493034698069096, + 0.005409623496234417, + 0.00413507828488946, + 0.0037944905925542116, + 0.003181505249813199, + 0.006023569032549858, + 0.007587619125843048, + 0.00784588884562254, + 0.012796035036444664, + 0.006410247180610895, + 0.005606678780168295, + 0.008009453304111958, + 0.006467197556048632, + 0.0076677026227116585, + 0.0021255947649478912, + 0.001270826323889196, + 0.0023851299192756414, + 0.001118950662203133, + 0.0014432996977120638, + 0.0015666330000385642, + 0.0013988062273710966, + 0.003968630917370319, + 0.0020500654354691505, + 0.0028575395699590445, + 0.0024816240184009075, + 0.0015131068648770452, + 0.001720102853141725, + 0.00170518783852458, + 0.0018116317223757505, + 0.0039057473186403513, + 0.004004006739705801, + 0.0038225743919610977, + 0.0048305620439350605, + 0.0023479026276618242, + 0.007504114881157875, + 0.004372460301965475, + 0.0034006370697170496, + 0.003044681390747428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.13644374907016754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636779099702835 + ], + "opposite_logprob_tokens": [ + -3.6513679027557373 + ], + "logprob_reward": -0.02636779099702835, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636779099702835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2101033478975296, + 0.08046364039182663, + 0.09239700436592102, + 0.031155528500676155, + 0.10466640442609787, + 0.05844122916460037, + 0.2096247673034668, + 0.055393632501363754, + 0.03787533938884735, + 0.11987914890050888 + ], + "credit_map": [ + 0.008840801194310188, + 0.01225166954100132, + 0.019415276125073433, + 0.008960551582276821, + 0.015940498560667038, + 0.07272276282310486, + 0.018384220078587532, + 0.05358756706118584, + 0.010516482405364513, + 0.012588872574269772, + 0.011315860785543919, + 0.00868910364806652, + 0.007407532073557377, + 0.012598537839949131, + 0.006820902228355408, + 0.010526352562010288, + 0.008984414860606194, + 0.008517109788954258, + 0.012171762064099312, + 0.008896770887076855, + 0.011818582192063332, + 0.010824094526469707, + 0.01745772361755371, + 0.013726540841162205, + 0.004519697278738022, + 0.0037593001034110785, + 0.00438164034858346, + 0.0037660172674804926, + 0.0024171979166567326, + 0.004077120218425989, + 0.0033078258857131004, + 0.004926727153360844, + 0.01109340414404869, + 0.014724353328347206, + 0.01591263711452484, + 0.0119996452704072, + 0.011225729249417782, + 0.012042048387229443, + 0.011478470638394356, + 0.01619010977447033, + 0.007038932293653488, + 0.0055487072095274925, + 0.008206816390156746, + 0.007744453847408295, + 0.005423408001661301, + 0.007594382390379906, + 0.00497250072658062, + 0.011912023648619652, + 0.027113601565361023, + 0.018049443140625954, + 0.04142841696739197, + 0.022460410371422768, + 0.01636858657002449, + 0.019021887332201004, + 0.027847204357385635, + 0.03733520582318306, + 0.0062027499079704285, + 0.006509428843855858, + 0.00953046977519989, + 0.005937402136623859, + 0.006058304104954004, + 0.006520450115203857, + 0.004711132030934095, + 0.009923691861331463, + 0.003942747600376606, + 0.005421174690127373, + 0.004086380824446678, + 0.0031154844909906387, + 0.0033610155805945396, + 0.0038702008314430714, + 0.006246779579669237, + 0.007831552997231483, + 0.015505180694162846, + 0.010522397235035896, + 0.02413857728242874, + 0.008184336125850677, + 0.021204253658652306, + 0.01377799641340971, + 0.010608362965285778, + 0.015938038006424904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.02636779099702835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005644158809445798 + ], + "opposite_logprob_tokens": [ + -7.5005645751953125 + ], + "logprob_reward": -0.0005644158809445798, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005644158809445798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8321043252944946, + 0.016849501058459282, + 0.04125260189175606, + 0.006825938355177641, + 0.028512394055724144, + 0.014313824474811554, + 0.018393728882074356, + 0.01979675516486168, + 0.014585007913410664, + 0.007365915458649397 + ], + "credit_map": [ + 0.03128044679760933, + 0.010710833594202995, + 0.08561651408672333, + 0.013555947691202164, + 0.053771935403347015, + 0.3714323043823242, + 0.018112093210220337, + 0.24762432277202606, + 0.0019682450219988823, + 0.0019225046271458268, + 0.0020501448307186365, + 0.0013622501865029335, + 0.00202493485994637, + 0.0029227337799966335, + 0.0016029935795813799, + 0.0029956954531371593, + 0.0055288951843976974, + 0.00546234892681241, + 0.0056736827827990055, + 0.003432970494031906, + 0.003459558356553316, + 0.004506301134824753, + 0.0069281840696930885, + 0.0062606665305793285, + 0.0009126257500611246, + 0.0006511080428026617, + 0.0012048023054376245, + 0.0007909155683591962, + 0.0006531514809466898, + 0.0009054343681782484, + 0.0007084857788868248, + 0.0009994151769205928, + 0.0034701756667345762, + 0.0031514805741608143, + 0.0046812500804662704, + 0.0030052487272769213, + 0.0029109674505889416, + 0.0033906488679349422, + 0.0028429648373275995, + 0.0050596618093550205, + 0.0017570179188624024, + 0.001423757872544229, + 0.0020095137879252434, + 0.0019722634460777044, + 0.0015236760955303907, + 0.0013634231872856617, + 0.0012209316482767463, + 0.003043241100385785, + 0.002819050569087267, + 0.0015159890754148364, + 0.003474789671599865, + 0.0017214741092175245, + 0.0017120575066655874, + 0.0020731494296342134, + 0.002080702455714345, + 0.0029965194407850504, + 0.002475939691066742, + 0.0019448264501988888, + 0.0024369757156819105, + 0.0015164061915129423, + 0.0015222199726849794, + 0.0022962461225688457, + 0.0021888986229896545, + 0.005415243562310934, + 0.001798021956346929, + 0.0018683011876419187, + 0.0018630854319781065, + 0.0013926364481449127, + 0.001424618298187852, + 0.0016148420982062817, + 0.002415608149021864, + 0.0022078948095440865, + 0.0009533298434689641, + 0.0006534048588946462, + 0.0011660457821562886, + 0.0005327326944097877, + 0.001089436816982925, + 0.0010271065402776003, + 0.0005842232494615018, + 0.0013596363132819533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0005644158809445798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603543907403946 + ], + "opposite_logprob_tokens": [ + -1.9103543758392334 + ], + "logprob_reward": -0.1603543907403946, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603543907403946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7503764033317566, + 0.04183896631002426, + 0.027737189084291458, + 0.017306100577116013, + 0.03808007016777992, + 0.024379504844546318, + 0.040340159088373184, + 0.01770351268351078, + 0.013202455826103687, + 0.02903561480343342 + ], + "credit_map": [ + 0.027620481327176094, + 0.01281589176505804, + 0.07466293126344681, + 0.013527298346161842, + 0.04401695728302002, + 0.3297571837902069, + 0.015564137138426304, + 0.23241156339645386, + 0.004948433954268694, + 0.004060686100274324, + 0.004563933238387108, + 0.003508045570924878, + 0.00377565692178905, + 0.007771568838506937, + 0.003430027747526765, + 0.009780612774193287, + 0.003864019876345992, + 0.0036504887975752354, + 0.003544932696968317, + 0.002692303853109479, + 0.0031618045177310705, + 0.0033095157705247402, + 0.003642162075266242, + 0.003871960798278451, + 0.002812282182276249, + 0.001645806827582419, + 0.0020906589925289154, + 0.0013350513763725758, + 0.0011991949286311865, + 0.0029374202713370323, + 0.0011751777492463589, + 0.004110509995371103, + 0.007201579865068197, + 0.0025555333122611046, + 0.005750983487814665, + 0.0033962558954954147, + 0.0035532547626644373, + 0.004960062447935343, + 0.002694436116144061, + 0.007967964746057987, + 0.0029813707806169987, + 0.002701098332181573, + 0.0033119062427431345, + 0.002889408031478524, + 0.0025186841376125813, + 0.0030486227478832006, + 0.0023483778350055218, + 0.004580034874379635, + 0.004770071245729923, + 0.004630511160939932, + 0.008819874376058578, + 0.004254479892551899, + 0.003736344864591956, + 0.00473660696297884, + 0.004647487308830023, + 0.004744785372167826, + 0.002406894229352474, + 0.0019372737733647227, + 0.0026209568604826927, + 0.0015154009452089667, + 0.0021407250314950943, + 0.0018523179460316896, + 0.0016358172288164496, + 0.003594125621020794, + 0.0016417806036770344, + 0.0018592745764181018, + 0.0017539073014631867, + 0.000821337744127959, + 0.0010536233894526958, + 0.0014618303393945098, + 0.001346084987744689, + 0.0032646171748638153, + 0.0030812486074864864, + 0.0024952245876193047, + 0.004461273085325956, + 0.0020376541651785374, + 0.006083692889660597, + 0.003452736185863614, + 0.0028928290121257305, + 0.004530956503003836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.1603543907403946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 close-up that circles from left to right around the fireplace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several stout birch logs supporting the fire, with charred black surfaces and visible fiery red cores?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do brilliant orange-yellow flames dance and flicker, punctuated by golden sparks drifting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the hearth made of rough, unpolished natural stone with a cast iron interior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the flames serve as the sole primary light source, casting warm, flickering shadows on the hearth and stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097993951290846 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097993951290846, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097993951290846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4629227817058563, + 0.08084725588560104, + 0.03666500374674797, + 0.09835519641637802, + 0.04924783110618591, + 0.09899158030748367, + 0.027158627286553383, + 0.03488534688949585, + 0.03530450910329819, + 0.07562179863452911 + ], + "credit_map": [ + 0.02106262370944023, + 0.01112416759133339, + 0.04998650774359703, + 0.01087066251784563, + 0.029682068154215813, + 0.19664476811885834, + 0.01699775829911232, + 0.1265542358160019, + 0.011546504683792591, + 0.01347285695374012, + 0.009174488484859467, + 0.008326707407832146, + 0.008041379041969776, + 0.01107658352702856, + 0.006768636405467987, + 0.0124400919303298, + 0.003960882313549519, + 0.0031565632671117783, + 0.0045992592349648476, + 0.0032309761736541986, + 0.0039826445281505585, + 0.004194148816168308, + 0.0064692310988903046, + 0.007071296218782663, + 0.014569195918738842, + 0.01357281394302845, + 0.011440293863415718, + 0.00647350400686264, + 0.007630201987922192, + 0.015181947499513626, + 0.007437857333570719, + 0.022049376741051674, + 0.008180584758520126, + 0.0062473490834236145, + 0.007613760884851217, + 0.0036829886958003044, + 0.005445888265967369, + 0.0057778977788984776, + 0.004495788831263781, + 0.0078035760670900345, + 0.01406630128622055, + 0.0093141570687294, + 0.013871515169739723, + 0.012239058502018452, + 0.009996081702411175, + 0.009976837784051895, + 0.00853646732866764, + 0.020991167053580284, + 0.003415679791942239, + 0.0041433884762227535, + 0.004614570643752813, + 0.0022890137042850256, + 0.002774004591628909, + 0.003531790804117918, + 0.0022693697828799486, + 0.004120809491723776, + 0.0038749580271542072, + 0.004283485934138298, + 0.003944004885852337, + 0.00277960067614913, + 0.003168742172420025, + 0.0035987519659101963, + 0.005069093778729439, + 0.008166709914803505, + 0.004548557102680206, + 0.00460575008764863, + 0.0037485267966985703, + 0.0025421571917831898, + 0.0027779012452811003, + 0.004007084295153618, + 0.005361835937947035, + 0.0077126952819526196, + 0.011146850883960724, + 0.007726595737040043, + 0.014091856777668, + 0.005356261041015387, + 0.013453966937959194, + 0.009048866108059883, + 0.0074245510622859, + 0.0073728496208786964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.004097993951290846 + } + ], + "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/0000131/group_5/index_tensor([3825, 3825])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_5/index_tensor([3825, 3825])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..893b20b614981a3d64d049b6756b2933cb3f025c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_5/index_tensor([3825, 3825])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning low-angle shot, the camera practically clinging to the rain-soaked black asphalt, zooms in with a violent, precarious urge toward the rear tire of a Lamborghini Aventador SVJ, which is furiously burning tires. Deep within a rainy cyberpunk metropolis, towering buildings cast vast shadows, their surfaces gleaming with giant holographic advertisements. Magenta, cyan, and electric blue neon lights stretch and twist across the wet road, forming streams of vibrant light. The angular Lamborghini Aventador SVJ, clad in matte black carbon fiber, its sharp lines and exaggerated aerodynamics create a cool silhouette against the shifting neon light. The absolute focal point of the shot—the rear wheel spinning at breakneck speed—vibrates violently from the immense friction, the wide, high-performance tires practically tearing apart, sending rubber particles whistling and scattering. Thick, scalding white smoke erupts from the point where the tires meet the ground, surging out like a volcanic eruption, quickly engulfing the space behind the car body and presenting psychedelic color layers under the reflection of neon lights. The entire picture is full of the ultimate details and dynamic tension of top cyberpunk animation, creating an adrenaline-pumping and shocking atmosphere of primitive power and future technology.", + "seed": 9256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_5/index_3825_seed_9256_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": -0.039363324735313655, + "vlm_logprob": -0.039363324735313655, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.039363324735313655, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06597632169723511 + ], + "opposite_logprob_tokens": [ + -3.81597638130188 + ], + "logprob_reward": -0.06597632169723511, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.06597632169723511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06915201246738434, + 0.03386455774307251, + 0.14305128157138824, + 0.04438299313187599, + 0.04774908348917961, + 0.05997664853930473, + 0.11993677914142609, + 0.030074162408709526, + 0.14101599156856537, + 0.31079643964767456 + ], + "credit_map": [ + 0.010522165335714817, + 0.005937238223850727, + 0.007830996997654438, + 0.011446175165474415, + 0.00705871544778347, + 0.008218438364565372, + 0.008731557056307793, + 0.00940672680735588, + 0.003439453197643161, + 0.005869822110980749, + 0.0062577263452112675, + 0.004454188514500856, + 0.003709631972014904, + 0.0030179969035089016, + 0.0030383262783288956, + 0.004077414516359568, + 0.012149931862950325, + 0.03354521840810776, + 0.007440974935889244, + 0.007400361821055412, + 0.009896856732666492, + 0.009237588383257389, + 0.04279962554574013, + 0.020580729469656944, + 0.010740947909653187, + 0.003755369922146201, + 0.0038176150992512703, + 0.0071325539611279964, + 0.004333624616265297, + 0.0037231065798550844, + 0.006347780581563711, + 0.004531996324658394, + 0.0034218216314911842, + 0.00432294188067317, + 0.004954836331307888, + 0.011528965085744858, + 0.004160393960773945, + 0.0027403156273066998, + 0.011403728276491165, + 0.005216086283326149, + 0.005132812540978193, + 0.009222045540809631, + 0.009939022362232208, + 0.010138697922229767, + 0.008497627452015877, + 0.0036210448015481234, + 0.007928626611828804, + 0.005496778059750795, + 0.008008476346731186, + 0.016340656206011772, + 0.0181902963668108, + 0.022699130699038506, + 0.014583437703549862, + 0.008370814844965935, + 0.020206447690725327, + 0.011537524871528149, + 0.0032710859086364508, + 0.0025483197532594204, + 0.0055428557097911835, + 0.0047748577781021595, + 0.0034197703935205936, + 0.0025311121717095375, + 0.004643471445888281, + 0.0033426927402615547, + 0.008261016570031643, + 0.010604716837406158, + 0.027682805433869362, + 0.030687836930155754, + 0.009605023078620434, + 0.0087861567735672, + 0.02335832454264164, + 0.022030122578144073, + 0.015441099181771278, + 0.013495958410203457, + 0.12539324164390564, + 0.10762687027454376, + 0.004621633794158697, + 0.01478438451886177, + 0.009528621099889278, + 0.019904661923646927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.06597632169723511 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907130360603333 + ], + "opposite_logprob_tokens": [ + -2.464071273803711 + ], + "logprob_reward": -0.08907130360603333, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907130360603333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04099693149328232, + 0.033018894493579865, + 0.4197731912136078, + 0.1096498891711235, + 0.07570703327655792, + 0.0845647007226944, + 0.05069812387228012, + 0.06034989655017853, + 0.04502128064632416, + 0.0802200436592102 + ], + "credit_map": [ + 0.004177185241132975, + 0.0030235699377954006, + 0.003382686758413911, + 0.006721778307110071, + 0.0039266180247068405, + 0.007386018056422472, + 0.006644994020462036, + 0.005734080448746681, + 0.003083521965891123, + 0.003068377962335944, + 0.0046361698769032955, + 0.005202526226639748, + 0.004629321396350861, + 0.004232593346387148, + 0.0034045083448290825, + 0.00476187514141202, + 0.03425281494855881, + 0.10083674639463425, + 0.017050309106707573, + 0.021548384800553322, + 0.02981623075902462, + 0.024283628910779953, + 0.13055874407291412, + 0.06142636016011238, + 0.011489290744066238, + 0.01173926331102848, + 0.011098603717982769, + 0.019240863621234894, + 0.013655276037752628, + 0.01484640222042799, + 0.013675689697265625, + 0.01390449982136488, + 0.007043986581265926, + 0.009669873863458633, + 0.0064574540592730045, + 0.016615252941846848, + 0.007460766471922398, + 0.004561663139611483, + 0.01674921065568924, + 0.007148823235183954, + 0.004143096506595612, + 0.005630370695143938, + 0.01329885609447956, + 0.02211683988571167, + 0.006836838088929653, + 0.005190613679587841, + 0.01924060843884945, + 0.008107472211122513, + 0.003248604480177164, + 0.004704505670815706, + 0.010090513154864311, + 0.012064809910953045, + 0.00516556529328227, + 0.0041893054731190205, + 0.0065031275153160095, + 0.0047316886484622955, + 0.004072290379554033, + 0.006028106901794672, + 0.010491293855011463, + 0.008135168813169003, + 0.00594236608594656, + 0.006586750503629446, + 0.009765731170773506, + 0.00932819303125143, + 0.004313315264880657, + 0.007460381370037794, + 0.006871180143207312, + 0.00737896841019392, + 0.0037616605404764414, + 0.004266243893653154, + 0.006190294399857521, + 0.004779239185154438, + 0.003946807235479355, + 0.003909137565642595, + 0.030065571889281273, + 0.026015758514404297, + 0.002237565815448761, + 0.004605927038937807, + 0.004190940875560045, + 0.005248326808214188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.08907130360603333 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329203300178051 + ], + "opposite_logprob_tokens": [ + -3.773292064666748 + ], + "logprob_reward": -0.02329203300178051, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329203300178051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0340765118598938, + 0.012152671813964844, + 0.15227317810058594, + 0.02811289392411709, + 0.03694701939821243, + 0.03393595665693283, + 0.1904805600643158, + 0.0311657153069973, + 0.321084588766098, + 0.15977084636688232 + ], + "credit_map": [ + 0.0032683636527508497, + 0.0023385859094560146, + 0.00430133193731308, + 0.00609135627746582, + 0.004139590542763472, + 0.004653586074709892, + 0.0049785589799284935, + 0.004305142443627119, + 0.0011987964389845729, + 0.002222067443653941, + 0.00215072906576097, + 0.0013059721095487475, + 0.0013264447916299105, + 0.0013618136290460825, + 0.0010230881161987782, + 0.0015637606848031282, + 0.012484067119657993, + 0.038144201040267944, + 0.0067111775279045105, + 0.007914917543530464, + 0.01069420762360096, + 0.008689437992870808, + 0.04604794830083847, + 0.021587228402495384, + 0.005806459113955498, + 0.002402527956292033, + 0.0026396182365715504, + 0.004613547120243311, + 0.0027961083687841892, + 0.00316300755366683, + 0.003724627662450075, + 0.002966999774798751, + 0.0029856055043637753, + 0.0034385817125439644, + 0.003721520770341158, + 0.008431465364992619, + 0.004487059079110622, + 0.0022146005649119616, + 0.0069221006706357, + 0.004746088758111, + 0.0023788795806467533, + 0.0022984298411756754, + 0.003019573399797082, + 0.01104725431650877, + 0.0026307697407901287, + 0.001733100158162415, + 0.00871133804321289, + 0.002116613555699587, + 0.00921297911554575, + 0.011013583280146122, + 0.05879105255007744, + 0.06679821014404297, + 0.009232801385223866, + 0.008698703721165657, + 0.012819753959774971, + 0.01391349732875824, + 0.002003608737140894, + 0.0029747732914984226, + 0.005645768251270056, + 0.004122827667742968, + 0.0026362119242548943, + 0.002833850681781769, + 0.005117511842399836, + 0.005831162910908461, + 0.014058474451303482, + 0.01951618678867817, + 0.06926462054252625, + 0.07716279476881027, + 0.022952234372496605, + 0.01704934611916542, + 0.04888438805937767, + 0.05219655856490135, + 0.007198152132332325, + 0.005415867082774639, + 0.06560660153627396, + 0.05675876885652542, + 0.0028316467069089413, + 0.007092065643519163, + 0.005139266140758991, + 0.009728487581014633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.02329203300178051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594549672678113 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.00594549672678113, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594549672678113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07094969600439072, + 0.22580847144126892, + 0.20782238245010376, + 0.05253665894269943, + 0.023872798308730125, + 0.036967191845178604, + 0.0487639494240284, + 0.055146943777799606, + 0.12245053797960281, + 0.15568140149116516 + ], + "credit_map": [ + 0.006656458601355553, + 0.005522871855646372, + 0.006193347740918398, + 0.009445082396268845, + 0.006556938402354717, + 0.01838923804461956, + 0.008579183369874954, + 0.009606573730707169, + 0.018681680783629417, + 0.05415919050574303, + 0.06259443610906601, + 0.02186933346092701, + 0.02128976583480835, + 0.016220221295952797, + 0.012087580747902393, + 0.018906259909272194, + 0.01897033490240574, + 0.04316221550107002, + 0.014751174487173557, + 0.010433756746351719, + 0.014444581232964993, + 0.014190329238772392, + 0.0620741993188858, + 0.02979578636586666, + 0.00804710853844881, + 0.004640139639377594, + 0.0049796937964856625, + 0.00842523854225874, + 0.006393532268702984, + 0.006893675774335861, + 0.007126152981072664, + 0.006031116936355829, + 0.002421770943328738, + 0.0037203466054052114, + 0.002132310066372156, + 0.004322865046560764, + 0.003410827135667205, + 0.0015583522617816925, + 0.003429458709433675, + 0.0028768654447048903, + 0.003058781148865819, + 0.005186845548450947, + 0.005667800549417734, + 0.00605700584128499, + 0.005987765733152628, + 0.002648472785949707, + 0.004808324854820967, + 0.003552195616066456, + 0.003460499458014965, + 0.0033560972660779953, + 0.010840116068720818, + 0.013365624472498894, + 0.0035740649327635765, + 0.0034394152462482452, + 0.005198369733989239, + 0.005529760383069515, + 0.0035437708720564842, + 0.0065917037427425385, + 0.009732330217957497, + 0.007763213012367487, + 0.005369142163544893, + 0.0061098737642169, + 0.007924306206405163, + 0.008112600073218346, + 0.004951357375830412, + 0.006817995570600033, + 0.03098936378955841, + 0.03409748524427414, + 0.007305386010557413, + 0.005720389541238546, + 0.01761649362742901, + 0.014952072873711586, + 0.0080681461840868, + 0.007911733351647854, + 0.05369099602103233, + 0.04845069721341133, + 0.007135648746043444, + 0.009563015773892403, + 0.012539410963654518, + 0.008321750909090042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.00594549672678113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531468644738197 + ], + "opposite_logprob_tokens": [ + -4.387531280517578 + ], + "logprob_reward": -0.012531468644738197, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531468644738197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13659721612930298, + 0.02938980609178543, + 0.11756081134080887, + 0.12182249128818512, + 0.041932363063097, + 0.035919222980737686, + 0.06189195439219475, + 0.06399668008089066, + 0.030257171019911766, + 0.3606322705745697 + ], + "credit_map": [ + 0.018108783289790154, + 0.012587104924023151, + 0.017001943662762642, + 0.02023601345717907, + 0.014584236778318882, + 0.01871749572455883, + 0.020822374150156975, + 0.014539270661771297, + 0.0027856165543198586, + 0.002946126740425825, + 0.0035517876967787743, + 0.0041313269175589085, + 0.0031007654033601284, + 0.003968516364693642, + 0.0038824069779366255, + 0.005023259669542313, + 0.020020034164190292, + 0.008861714042723179, + 0.024946020916104317, + 0.009730798192322254, + 0.008644535206258297, + 0.009502017870545387, + 0.023532569408416748, + 0.012323125265538692, + 0.019576184451580048, + 0.011114018969237804, + 0.01854117400944233, + 0.01441881898790598, + 0.012729118578135967, + 0.013616867363452911, + 0.016204385086894035, + 0.01562192477285862, + 0.006155666895210743, + 0.00469624949619174, + 0.004082247614860535, + 0.007209497503936291, + 0.00533960061147809, + 0.0030118287540972233, + 0.0069225626066327095, + 0.0045147063210606575, + 0.002078533172607422, + 0.0025746640749275684, + 0.003065321361646056, + 0.011589410714805126, + 0.0028253609780222178, + 0.0019601185340434313, + 0.008920843712985516, + 0.0029049699660390615, + 0.0033616703003644943, + 0.005260997451841831, + 0.015187927521765232, + 0.018041463568806648, + 0.004980925004929304, + 0.0037429947406053543, + 0.006904108449816704, + 0.004411863628774881, + 0.0037017925642430782, + 0.007044801488518715, + 0.011693685315549374, + 0.008469789288938046, + 0.006191439926624298, + 0.006244117394089699, + 0.010159256868064404, + 0.010491794906556606, + 0.0019509535050019622, + 0.0028479350730776787, + 0.005070550832897425, + 0.005868243984878063, + 0.002409379929304123, + 0.0023286191280931234, + 0.004883423447608948, + 0.004898064769804478, + 0.01675509847700596, + 0.012679447419941425, + 0.14898161590099335, + 0.12875722348690033, + 0.0036085587926208973, + 0.016227848827838898, + 0.01027686893939972, + 0.02334558591246605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.012531468644738197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle shot, the camera practically clinging to the rain-soaked black asphalt, zooms in with a violent, precarious urge toward the rear tire of a Lamborghini Aventador SVJ, which is furiously burning tires. Deep within a rainy cyberpunk metropolis, towering buildings cast vast shadows, their surfaces gleaming with giant holographic advertisements. Magenta, cyan, and electric blue neon lights stretch and twist across the wet road, forming streams of vibrant light. The angular Lamborghini Aventador SVJ, clad in matte black carbon fiber, its sharp lines and exaggerated aerodynamics create a cool silhouette against the shifting neon light. The absolute focal point of the shot—the rear wheel spinning at breakneck speed—vibrates violently from the immense friction, the wide, high-performance tires practically tearing apart, sending rubber particles whistling and scattering. Thick, scalding white smoke erupts from the point where the tires meet the ground, surging out like a volcanic eruption, quickly engulfing the space behind the car body and presenting psychedelic color layers under the reflection of neon lights. The entire picture is full of the ultimate details and dynamic tension of top cyberpunk animation, creating an adrenaline-pumping and shocking atmosphere of primitive power and future technology.", + "seed": 9257, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_5/index_3825_seed_9257_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": -0.038419552613049746, + "vlm_logprob": -0.038419552613049746, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.038419552613049746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06415148079395294 + ], + "opposite_logprob_tokens": [ + -3.3141515254974365 + ], + "logprob_reward": -0.06415148079395294, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06415148079395294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0884518101811409, + 0.0833461582660675, + 0.17274563014507294, + 0.09845682233572006, + 0.08554843813180923, + 0.12943002581596375, + 0.05366380885243416, + 0.026551488786935806, + 0.16208510100841522, + 0.09972074627876282 + ], + "credit_map": [ + 0.010877266526222229, + 0.009659630246460438, + 0.009559141471982002, + 0.013472747057676315, + 0.008534972555935383, + 0.010124733671545982, + 0.013754907995462418, + 0.012468401342630386, + 0.006351645104587078, + 0.011685463599860668, + 0.0077568464912474155, + 0.01550218928605318, + 0.009835432283580303, + 0.010357112623751163, + 0.013203122653067112, + 0.008654339239001274, + 0.010415631346404552, + 0.014186268672347069, + 0.030517803505063057, + 0.01866443082690239, + 0.012835955247282982, + 0.007646792568266392, + 0.04679575935006142, + 0.03168298304080963, + 0.011116658337414265, + 0.00785077828913927, + 0.016274116933345795, + 0.01425752229988575, + 0.007933412678539753, + 0.009488444775342941, + 0.0199488066136837, + 0.011587073095142841, + 0.007158280815929174, + 0.008186182007193565, + 0.0069455537013709545, + 0.012892664410173893, + 0.013455936685204506, + 0.01016838476061821, + 0.01522384025156498, + 0.011517586186528206, + 0.011399442330002785, + 0.009333111345767975, + 0.021677259355783463, + 0.03517135977745056, + 0.008658270351588726, + 0.005418654531240463, + 0.03194443881511688, + 0.005827473476529121, + 0.0034227475989609957, + 0.0032354153227061033, + 0.004220287781208754, + 0.01621568202972412, + 0.004479410592466593, + 0.0033210599794983864, + 0.015510201454162598, + 0.003259001299738884, + 0.0033248693216592073, + 0.002492505358532071, + 0.0031607497949153185, + 0.0036457993555814028, + 0.0030414569191634655, + 0.002570437965914607, + 0.004437011666595936, + 0.0038786563090980053, + 0.020599864423274994, + 0.010411479510366917, + 0.03590139374136925, + 0.023965241387486458, + 0.010420789942145348, + 0.011273495852947235, + 0.03833821043372154, + 0.011174604296684265, + 0.022291172295808792, + 0.009119455702602863, + 0.011610547080636024, + 0.01764475367963314, + 0.0078042917884886265, + 0.004473073873668909, + 0.01319955289363861, + 0.013577892445027828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.06415148079395294 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005227357149124 + ], + "opposite_logprob_tokens": [ + -2.6950523853302 + ], + "logprob_reward": -0.07005227357149124, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005227357149124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0405898317694664, + 0.06231791153550148, + 0.16629186272621155, + 0.04606630280613899, + 0.2383953183889389, + 0.17068463563919067, + 0.14631320536136627, + 0.024436691775918007, + 0.06338024139404297, + 0.0415240041911602 + ], + "credit_map": [ + 0.00400913180783391, + 0.005560615565627813, + 0.005401402711868286, + 0.007005239371210337, + 0.0032909405417740345, + 0.0036826266441494226, + 0.007090908940881491, + 0.004548965021967888, + 0.005841396749019623, + 0.012194852344691753, + 0.00708013866096735, + 0.008115850389003754, + 0.004922740161418915, + 0.005252233240753412, + 0.010859334841370583, + 0.008051363751292229, + 0.01110083982348442, + 0.009167972020804882, + 0.030101891607046127, + 0.016416577622294426, + 0.008876997977495193, + 0.005598563235253096, + 0.04967593029141426, + 0.035353079438209534, + 0.0044744666665792465, + 0.0030348021537065506, + 0.006845959462225437, + 0.007920553907752037, + 0.004323418717831373, + 0.0033067709300667048, + 0.00983467884361744, + 0.006325653288513422, + 0.01250145211815834, + 0.00955123733729124, + 0.03756373003125191, + 0.04927202686667442, + 0.016366226598620415, + 0.015111256390810013, + 0.08249948173761368, + 0.015529919415712357, + 0.018874939531087875, + 0.019681937992572784, + 0.020582206547260284, + 0.024211803451180458, + 0.02596239373087883, + 0.017636865377426147, + 0.02420874685049057, + 0.019525747746229172, + 0.015672920271754265, + 0.01325348112732172, + 0.024500245228409767, + 0.02399527095258236, + 0.017690137028694153, + 0.012550003826618195, + 0.02532866783440113, + 0.013322485610842705, + 0.0030229135882109404, + 0.0031120481435209513, + 0.0027459440752863884, + 0.002707570791244507, + 0.0035544002894312143, + 0.0028325594030320644, + 0.0030488113407045603, + 0.0034124429803341627, + 0.0046827238984405994, + 0.00493780430406332, + 0.013713008724153042, + 0.008853321895003319, + 0.004837536718696356, + 0.003519330406561494, + 0.018668605014681816, + 0.0041679092682898045, + 0.00742902047932148, + 0.005089413840323687, + 0.005320250056684017, + 0.005534050054848194, + 0.002693277783691883, + 0.0025968002155423164, + 0.007394108921289444, + 0.005467085633426905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.07005227357149124 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590132102370262 + ], + "opposite_logprob_tokens": [ + -3.89559006690979 + ], + "logprob_reward": -0.020590132102370262, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590132102370262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07628653943538666, + 0.2111925482749939, + 0.13190490007400513, + 0.0628724917769432, + 0.14274844527244568, + 0.09027338773012161, + 0.092945896089077, + 0.04824799299240112, + 0.08099365234375, + 0.06253419816493988 + ], + "credit_map": [ + 0.006837063003331423, + 0.006587542127817869, + 0.007484858389943838, + 0.02103857882320881, + 0.0041823359206318855, + 0.00486596766859293, + 0.019123176112771034, + 0.0061670150607824326, + 0.028087815269827843, + 0.05161803960800171, + 0.021550392732024193, + 0.03386324271559715, + 0.016247699037194252, + 0.00923515111207962, + 0.03598324581980705, + 0.01460695918649435, + 0.00772181386128068, + 0.008377647027373314, + 0.01902257278561592, + 0.01289869099855423, + 0.009067348204553127, + 0.005431578494608402, + 0.03970235213637352, + 0.029682882130146027, + 0.007425127550959587, + 0.004410740453749895, + 0.011219678446650505, + 0.00814092718064785, + 0.005653020925819874, + 0.0045026452280581, + 0.01194086391478777, + 0.009579481557011604, + 0.00782325305044651, + 0.009984324686229229, + 0.011056075803935528, + 0.024271782487630844, + 0.012347239069640636, + 0.018965302035212517, + 0.03710060194134712, + 0.02119986154139042, + 0.010542507283389568, + 0.006232129875570536, + 0.008545205928385258, + 0.01993325538933277, + 0.010877950116991997, + 0.005822927225381136, + 0.020077604800462723, + 0.008241811767220497, + 0.009963687509298325, + 0.009071915410459042, + 0.013150878250598907, + 0.015630178153514862, + 0.013257191516458988, + 0.007633754517883062, + 0.015773242339491844, + 0.008465041406452656, + 0.004912164062261581, + 0.005913218948990107, + 0.005750315263867378, + 0.008239329792559147, + 0.0038194044027477503, + 0.0037225752603262663, + 0.00967888068407774, + 0.0062121013179421425, + 0.00814796332269907, + 0.004321226384490728, + 0.018702995032072067, + 0.014335932210087776, + 0.0043144216760993, + 0.004503368865698576, + 0.021628374233841896, + 0.005039365030825138, + 0.013352741487324238, + 0.006572485901415348, + 0.005943041294813156, + 0.008685638196766376, + 0.00435061939060688, + 0.002925460459664464, + 0.010571903549134731, + 0.010132298804819584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.020590132102370262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04486068710684776, + 0.1240551695227623, + 0.22498001158237457, + 0.12587352097034454, + 0.12267137318849564, + 0.13894139230251312, + 0.06658361852169037, + 0.024724598973989487, + 0.045893456786870956, + 0.0814160704612732 + ], + "credit_map": [ + 0.005359196569770575, + 0.004801283124834299, + 0.005342354066669941, + 0.007445407100021839, + 0.0038084834814071655, + 0.004179401323199272, + 0.00732614379376173, + 0.006598416715860367, + 0.01131671667098999, + 0.03291836753487587, + 0.008672921918332577, + 0.0223858579993248, + 0.009349391795694828, + 0.0075928508304059505, + 0.02307193912565708, + 0.008747126907110214, + 0.01404170598834753, + 0.009043878875672817, + 0.04584630951285362, + 0.020113976672291756, + 0.012708541937172413, + 0.00798857118934393, + 0.06838821619749069, + 0.04684881493449211, + 0.01353245135396719, + 0.008514342829585075, + 0.022702345624566078, + 0.02254304848611355, + 0.010744868777692318, + 0.009531660005450249, + 0.02081727609038353, + 0.017487524077296257, + 0.008369402959942818, + 0.006631459575146437, + 0.017874276265501976, + 0.02359626069664955, + 0.00815578643232584, + 0.007200310006737709, + 0.0393885001540184, + 0.011455371975898743, + 0.009118383750319481, + 0.009104589000344276, + 0.01214179489761591, + 0.03775345906615257, + 0.018683357164263725, + 0.008329755626618862, + 0.035806696861982346, + 0.00800334569066763, + 0.006005868315696716, + 0.004762422293424606, + 0.006751343607902527, + 0.01381215825676918, + 0.007684440352022648, + 0.007470041513442993, + 0.014656389132142067, + 0.005440962500870228, + 0.0033353932667523623, + 0.002710541244596243, + 0.0031672946643084288, + 0.003356935689225793, + 0.0026094275526702404, + 0.0027109268121421337, + 0.003548380685970187, + 0.0032856985926628113, + 0.0039822692051529884, + 0.003516070544719696, + 0.009521746076643467, + 0.006820476148277521, + 0.0032913656905293465, + 0.0024044779129326344, + 0.012778324075043201, + 0.0035787264350801706, + 0.012457500211894512, + 0.006368555128574371, + 0.01114470511674881, + 0.014836365357041359, + 0.007060866337269545, + 0.005873281974345446, + 0.013472088612616062, + 0.010202708654105663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.0036137294955551624 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369014710187912 + ], + "opposite_logprob_tokens": [ + -3.4086902141571045 + ], + "logprob_reward": -0.03369014710187912, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369014710187912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10298691689968109, + 0.04238535091280937, + 0.16766607761383057, + 0.07797639816999435, + 0.3019099235534668, + 0.17266540229320526, + 0.01580091379582882, + 0.015513801947236061, + 0.03830387443304062, + 0.06479130685329437 + ], + "credit_map": [ + 0.01322803646326065, + 0.011046956293284893, + 0.01184612326323986, + 0.015207392163574696, + 0.009291533380746841, + 0.009898489341139793, + 0.01639474742114544, + 0.016073646023869514, + 0.00384170631878078, + 0.008126342669129372, + 0.0033689793199300766, + 0.005877855233848095, + 0.004636464174836874, + 0.00334224384278059, + 0.0072230384685099125, + 0.005968726705759764, + 0.010350797325372696, + 0.010473037138581276, + 0.030686693266034126, + 0.016271276399493217, + 0.009805647656321526, + 0.005932202097028494, + 0.04931892827153206, + 0.034827496856451035, + 0.009516195394098759, + 0.004774213768541813, + 0.013356586918234825, + 0.010781150311231613, + 0.007375073619186878, + 0.005415358114987612, + 0.015760526061058044, + 0.010997303761541843, + 0.019629111513495445, + 0.014458809979259968, + 0.056010644882917404, + 0.0579039603471756, + 0.01896536909043789, + 0.016950421035289764, + 0.0995863750576973, + 0.018405256792902946, + 0.022943971678614616, + 0.016100311651825905, + 0.033954307436943054, + 0.03321211412549019, + 0.014059825800359249, + 0.012218506075441837, + 0.028302380815148354, + 0.011874000541865826, + 0.0018262517405673862, + 0.0009207530529238284, + 0.0025264478754252195, + 0.0033604924101382494, + 0.0010754226241260767, + 0.0013255337253212929, + 0.0033537133131176233, + 0.0014122993452474475, + 0.0018485630862414837, + 0.0014328057877719402, + 0.002233353676274419, + 0.001891588093712926, + 0.0016918967012315989, + 0.0013371554668992758, + 0.003231718670576811, + 0.001846720464527607, + 0.0036223409697413445, + 0.003037823364138603, + 0.007058367133140564, + 0.00581403961405158, + 0.0031274110078811646, + 0.0024579388555139303, + 0.009836405515670776, + 0.0033495547249913216, + 0.01331263966858387, + 0.004262780770659447, + 0.008033343590795994, + 0.010345084592700005, + 0.0053757657296955585, + 0.004383232910186052, + 0.009716367349028587, + 0.009362093172967434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.03369014710187912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle shot, the camera practically clinging to the rain-soaked black asphalt, zooms in with a violent, precarious urge toward the rear tire of a Lamborghini Aventador SVJ, which is furiously burning tires. Deep within a rainy cyberpunk metropolis, towering buildings cast vast shadows, their surfaces gleaming with giant holographic advertisements. Magenta, cyan, and electric blue neon lights stretch and twist across the wet road, forming streams of vibrant light. The angular Lamborghini Aventador SVJ, clad in matte black carbon fiber, its sharp lines and exaggerated aerodynamics create a cool silhouette against the shifting neon light. The absolute focal point of the shot—the rear wheel spinning at breakneck speed—vibrates violently from the immense friction, the wide, high-performance tires practically tearing apart, sending rubber particles whistling and scattering. Thick, scalding white smoke erupts from the point where the tires meet the ground, surging out like a volcanic eruption, quickly engulfing the space behind the car body and presenting psychedelic color layers under the reflection of neon lights. The entire picture is full of the ultimate details and dynamic tension of top cyberpunk animation, creating an adrenaline-pumping and shocking atmosphere of primitive power and future technology.", + "seed": 9258, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_5/index_3825_seed_9258_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": -0.05363411251455545, + "vlm_logprob": -0.05363411251455545, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05363411251455545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.071550153195858 + ], + "opposite_logprob_tokens": [ + -3.321550130844116 + ], + "logprob_reward": -0.071550153195858, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.071550153195858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1977173537015915, + 0.06546356528997421, + 0.13920758664608002, + 0.09935542196035385, + 0.12289777398109436, + 0.07341223210096359, + 0.054801009595394135, + 0.15570324659347534, + 0.03495212644338608, + 0.056489650160074234 + ], + "credit_map": [ + 0.014538786374032497, + 0.01880374737083912, + 0.015256526879966259, + 0.0312860868871212, + 0.013073385693132877, + 0.0507102869451046, + 0.024832019582390785, + 0.029216516762971878, + 0.005678215529769659, + 0.005650176201015711, + 0.005306317936629057, + 0.013052371330559254, + 0.005855463445186615, + 0.00796969048678875, + 0.008480883203446865, + 0.013470450416207314, + 0.007075608242303133, + 0.00794227235019207, + 0.015472025610506535, + 0.022876080125570297, + 0.010315880179405212, + 0.011291232891380787, + 0.03367722034454346, + 0.030557261779904366, + 0.008049465715885162, + 0.011055230163037777, + 0.013649498112499714, + 0.020446984097361565, + 0.011517471633851528, + 0.009038802236318588, + 0.018093694001436234, + 0.007504284381866455, + 0.00894168671220541, + 0.012424428015947342, + 0.012053456157445908, + 0.014467881992459297, + 0.020775483921170235, + 0.016665784642100334, + 0.02101263962686062, + 0.01655641198158264, + 0.005154874641448259, + 0.006308958865702152, + 0.007733368314802647, + 0.008572643622756004, + 0.006401488557457924, + 0.015959542244672775, + 0.008947769179940224, + 0.01433358620852232, + 0.005168264266103506, + 0.005859158933162689, + 0.004533776547759771, + 0.009370077401399612, + 0.00563919497653842, + 0.005058288108557463, + 0.010345418937504292, + 0.008826827630400658, + 0.021511176601052284, + 0.013906455598771572, + 0.02725360356271267, + 0.02747599221765995, + 0.012069014832377434, + 0.012887579388916492, + 0.021314822137355804, + 0.019284604117274284, + 0.003698422107845545, + 0.0031640934757888317, + 0.004999119322746992, + 0.004984097555279732, + 0.003424108726903796, + 0.005371261853724718, + 0.005167006049305201, + 0.004144017584621906, + 0.005349408369511366, + 0.011019335128366947, + 0.005694386083632708, + 0.007254363503307104, + 0.005168417003005743, + 0.008224722929298878, + 0.006528353318572044, + 0.007250661496073008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.071550153195858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603897362947464 + ], + "opposite_logprob_tokens": [ + -1.91038978099823 + ], + "logprob_reward": -0.1603897362947464, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603897362947464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27333447337150574, + 0.083931103348732, + 0.15571971237659454, + 0.12381649017333984, + 0.034280382096767426, + 0.07787755131721497, + 0.10552561283111572, + 0.08218605071306229, + 0.0270224642008543, + 0.03630615398287773 + ], + "credit_map": [ + 0.01898167096078396, + 0.029407735913991928, + 0.014730338007211685, + 0.054465167224407196, + 0.013804805465042591, + 0.08730453997850418, + 0.022768575698137283, + 0.03187163174152374, + 0.005821585189551115, + 0.006049773655831814, + 0.008829336613416672, + 0.017452243715524673, + 0.008949767798185349, + 0.0069044786505401134, + 0.022532975301146507, + 0.007390942890197039, + 0.010470885783433914, + 0.007161946035921574, + 0.015051675960421562, + 0.02917046658694744, + 0.009161379188299179, + 0.012009389698505402, + 0.04350488632917404, + 0.029189083725214005, + 0.010554458945989609, + 0.010848701931536198, + 0.015600336715579033, + 0.023796657100319862, + 0.014339061453938484, + 0.009839960373938084, + 0.02940955013036728, + 0.009427770972251892, + 0.0046493918634951115, + 0.0036217549350112677, + 0.004444668535143137, + 0.0038954669144004583, + 0.00427160644903779, + 0.0038172975182533264, + 0.0061200885102152824, + 0.00346010853536427, + 0.010109685361385345, + 0.006278050132095814, + 0.011921132914721966, + 0.01263782661408186, + 0.007236642763018608, + 0.007997271604835987, + 0.010342552326619625, + 0.011354398913681507, + 0.010204412043094635, + 0.015812192112207413, + 0.01028463989496231, + 0.013147433288395405, + 0.013688390143215656, + 0.015659725293517113, + 0.016711244359612465, + 0.010017568245530128, + 0.006426383275538683, + 0.009573184885084629, + 0.007844432257115841, + 0.01592504233121872, + 0.008904491551220417, + 0.013012099079787731, + 0.011802810244262218, + 0.008697611279785633, + 0.004404653329402208, + 0.0032521935645490885, + 0.004255502950400114, + 0.00344310631044209, + 0.0023980119731277227, + 0.0026342193596065044, + 0.004022351000458002, + 0.002612426644191146, + 0.0035392488352954388, + 0.00550301605835557, + 0.007044036872684956, + 0.005331408232450485, + 0.00378982350230217, + 0.003068925580009818, + 0.003594856709241867, + 0.004434836562722921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1603897362947464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076814383268356 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.016076814383268356, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076814383268356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5840665102005005, + 0.07179035991430283, + 0.11255121976137161, + 0.05119413509964943, + 0.03413890674710274, + 0.02771652489900589, + 0.029556306079030037, + 0.04107086732983589, + 0.013166685588657856, + 0.03474847599864006 + ], + "credit_map": [ + 0.03822296857833862, + 0.05794892832636833, + 0.030147327110171318, + 0.1205325648188591, + 0.02795029804110527, + 0.1905081570148468, + 0.049003709107637405, + 0.06975258886814117, + 0.004732633009552956, + 0.006790094077587128, + 0.00835438072681427, + 0.015056741423904896, + 0.007300270721316338, + 0.005052805412560701, + 0.016263138502836227, + 0.008240297436714172, + 0.007786177098751068, + 0.007713336497545242, + 0.013005416840314865, + 0.023941954597830772, + 0.00760216498747468, + 0.006253664381802082, + 0.029627280309796333, + 0.01662122830748558, + 0.0051758745685219765, + 0.0047944881953299046, + 0.006785171572118998, + 0.011413446627557278, + 0.005610296968370676, + 0.004727092105895281, + 0.008269541896879673, + 0.004418219905346632, + 0.003437742590904236, + 0.0035570880863815546, + 0.002934681950137019, + 0.0036217104643583298, + 0.005336909554898739, + 0.006041096989065409, + 0.004908245522528887, + 0.004301430191844702, + 0.0025506920646876097, + 0.0029824464581906796, + 0.003038168651983142, + 0.004669898189604282, + 0.002755754394456744, + 0.003662481438368559, + 0.0035561486147344112, + 0.004500934388488531, + 0.003028670558705926, + 0.004014480393379927, + 0.003055989509448409, + 0.003419206477701664, + 0.0039189825765788555, + 0.004612149205058813, + 0.004027264658361673, + 0.003479562932625413, + 0.005127927754074335, + 0.003952941857278347, + 0.006546299904584885, + 0.007316936273127794, + 0.004549707286059856, + 0.0032486291602253914, + 0.006056478712707758, + 0.004271945916116238, + 0.0014395100297406316, + 0.0011749599361792207, + 0.00203673611395061, + 0.0017732620472088456, + 0.0014279456809163094, + 0.001612122287042439, + 0.002052469179034233, + 0.0016496804310008883, + 0.002727924147620797, + 0.005191201809793711, + 0.004069504328072071, + 0.006376815494149923, + 0.003254212671890855, + 0.003301188349723816, + 0.005367292556911707, + 0.004460333846509457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016076814383268356 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092295654118061 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092295654118061, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092295654118061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7049688100814819, + 0.026147644966840744, + 0.09204374253749847, + 0.03484547883272171, + 0.01835201308131218, + 0.01591036655008793, + 0.02926778793334961, + 0.03739134222269058, + 0.01734207198023796, + 0.023730747401714325 + ], + "credit_map": [ + 0.048662975430488586, + 0.07094145566225052, + 0.0340735949575901, + 0.14502020180225372, + 0.03652483969926834, + 0.22455231845378876, + 0.0593186654150486, + 0.08587475121021271, + 0.0022622570395469666, + 0.0031952825374901295, + 0.002247254829853773, + 0.00468985503539443, + 0.0022028167732059956, + 0.0032792063429951668, + 0.0035505020059645176, + 0.0047204685397446156, + 0.004445766564458609, + 0.003928440157324076, + 0.006850463338196278, + 0.018768643960356712, + 0.005227837711572647, + 0.005857216659933329, + 0.028228139504790306, + 0.018737228587269783, + 0.0031254789792001247, + 0.0034887883812189102, + 0.0039865924045443535, + 0.007175496779382229, + 0.0034310745541006327, + 0.0028605444822460413, + 0.0076531521044671535, + 0.0031243523117154837, + 0.0027189201209694147, + 0.0024881833232939243, + 0.001866117469035089, + 0.00215330277569592, + 0.0023085547145456076, + 0.0018899741116911173, + 0.0024049619678407907, + 0.002521998481824994, + 0.0011342628858983517, + 0.0011229963274672627, + 0.0015654639573767781, + 0.0019535210449248552, + 0.0013317704433575273, + 0.0036043638829141855, + 0.0019876204896718264, + 0.0032103662379086018, + 0.0029622577130794525, + 0.0020420884247869253, + 0.0026488674338907003, + 0.004325685556977987, + 0.0029631643556058407, + 0.003230615518987179, + 0.005399692803621292, + 0.00569541472941637, + 0.0045120189897716045, + 0.0027722029481083155, + 0.0077250306494534016, + 0.005021445918828249, + 0.0033362621907144785, + 0.0030596565920859575, + 0.0048613292165100574, + 0.0061033945530653, + 0.00166363175958395, + 0.002553457859903574, + 0.0025053848512470722, + 0.0018929161597043276, + 0.0014264184283092618, + 0.0025895254220813513, + 0.002568901050835848, + 0.0021418374963104725, + 0.002033430151641369, + 0.003436063649132848, + 0.0028583749663084745, + 0.004358434583991766, + 0.0021000048145651817, + 0.0020493913907557726, + 0.003663955954834819, + 0.003231092356145382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.004092295654118061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061563044786453 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061563044786453, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061563044786453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5674173831939697, + 0.05771314352750778, + 0.10873228311538696, + 0.07728014141321182, + 0.05282214283943176, + 0.01765393652021885, + 0.021430257707834244, + 0.05658605322241783, + 0.012025785632431507, + 0.028338918462395668 + ], + "credit_map": [ + 0.03808099776506424, + 0.05678221583366394, + 0.02770526520907879, + 0.1177438348531723, + 0.028024347499012947, + 0.18622823059558868, + 0.045803945511579514, + 0.06704846024513245, + 0.005834192968904972, + 0.004219279158860445, + 0.005314899608492851, + 0.011425582692027092, + 0.004949296824634075, + 0.0065396567806601524, + 0.009499005042016506, + 0.009931227192282677, + 0.00577534781768918, + 0.004183242563158274, + 0.006842279341071844, + 0.022834813222289085, + 0.0059480066411197186, + 0.006454951595515013, + 0.03391873836517334, + 0.022774899378418922, + 0.0077619352377951145, + 0.007125875446945429, + 0.014944762922823429, + 0.012455090880393982, + 0.006543386727571487, + 0.008652636781334877, + 0.012077638879418373, + 0.00771880429238081, + 0.005100504495203495, + 0.005283322650939226, + 0.005588212516158819, + 0.005134681239724159, + 0.007449763361364603, + 0.0124448137357831, + 0.005097868852317333, + 0.006722975522279739, + 0.0017360278870910406, + 0.001166133675724268, + 0.001804355881176889, + 0.0023236144334077835, + 0.001623094780370593, + 0.004298285115510225, + 0.0017718167509883642, + 0.00293060764670372, + 0.0021949114743620157, + 0.002745991572737694, + 0.0018310218583792448, + 0.003326654899865389, + 0.002592490054666996, + 0.0025805365294218063, + 0.003935644868761301, + 0.0022230050526559353, + 0.006916085723787546, + 0.004288079217076302, + 0.00870153121650219, + 0.01138305477797985, + 0.003735720179975033, + 0.004826724529266357, + 0.011018412187695503, + 0.005716440733522177, + 0.0014123886357992887, + 0.001281794044189155, + 0.0018389541655778885, + 0.001801160629838705, + 0.0012455362593755126, + 0.0018259957432746887, + 0.0017022405518218875, + 0.0009177137981168926, + 0.0037745151203125715, + 0.004506056196987629, + 0.0041944896802306175, + 0.003400924149900675, + 0.002922334708273411, + 0.002525283955037594, + 0.003358038142323494, + 0.003657273016870022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.016061563044786453 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle shot, the camera practically clinging to the rain-soaked black asphalt, zooms in with a violent, precarious urge toward the rear tire of a Lamborghini Aventador SVJ, which is furiously burning tires. Deep within a rainy cyberpunk metropolis, towering buildings cast vast shadows, their surfaces gleaming with giant holographic advertisements. Magenta, cyan, and electric blue neon lights stretch and twist across the wet road, forming streams of vibrant light. The angular Lamborghini Aventador SVJ, clad in matte black carbon fiber, its sharp lines and exaggerated aerodynamics create a cool silhouette against the shifting neon light. The absolute focal point of the shot—the rear wheel spinning at breakneck speed—vibrates violently from the immense friction, the wide, high-performance tires practically tearing apart, sending rubber particles whistling and scattering. Thick, scalding white smoke erupts from the point where the tires meet the ground, surging out like a volcanic eruption, quickly engulfing the space behind the car body and presenting psychedelic color layers under the reflection of neon lights. The entire picture is full of the ultimate details and dynamic tension of top cyberpunk animation, creating an adrenaline-pumping and shocking atmosphere of primitive power and future technology.", + "seed": 9259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_5/index_3825_seed_9259_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": -0.033640780672430995, + "vlm_logprob": -0.033640780672430995, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.033640780672430995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03190779313445091 + ], + "opposite_logprob_tokens": [ + -4.781907558441162 + ], + "logprob_reward": -0.03190779313445091, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.03190779313445091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07167304307222366, + 0.1285586953163147, + 0.04321284219622612, + 0.0621529184281826, + 0.05284851789474487, + 0.07840275019407272, + 0.05839333310723305, + 0.15126235783100128, + 0.2590588629245758, + 0.09443672001361847 + ], + "credit_map": [ + 0.005825169384479523, + 0.007979804649949074, + 0.009863302111625671, + 0.011572133749723434, + 0.005814079660922289, + 0.009105394594371319, + 0.014293651096522808, + 0.007219512015581131, + 0.008588429540395737, + 0.017681825906038284, + 0.011872487142682076, + 0.016177352517843246, + 0.018163707107305527, + 0.023008134216070175, + 0.022531278431415558, + 0.01053548138588667, + 0.0038893783930689096, + 0.006427191663533449, + 0.0049551124684512615, + 0.007414253428578377, + 0.004353223834186792, + 0.0028947852551937103, + 0.009498195722699165, + 0.0037807009648531675, + 0.0045793927274644375, + 0.007613564375787973, + 0.010852555744349957, + 0.011205339804291725, + 0.005485128611326218, + 0.003996501676738262, + 0.011911033652722836, + 0.006509402301162481, + 0.006149702705442905, + 0.004500444978475571, + 0.010670806281268597, + 0.011164483614265919, + 0.0037344947922974825, + 0.0027301341760903597, + 0.010743072256445885, + 0.003155376063659787, + 0.0071350811049342155, + 0.008698796853423119, + 0.008850150741636753, + 0.022795280441641808, + 0.003022801596671343, + 0.004868339281529188, + 0.018316447734832764, + 0.0047158533707261086, + 0.00895600114017725, + 0.00673655653372407, + 0.008336996659636497, + 0.00903236959129572, + 0.007008740678429604, + 0.006435469724237919, + 0.007184344343841076, + 0.004702855832874775, + 0.03461237996816635, + 0.016145072877407074, + 0.01752147264778614, + 0.01665942557156086, + 0.012270401231944561, + 0.01332328375428915, + 0.019727520644664764, + 0.021002812311053276, + 0.014849909581243992, + 0.02344542182981968, + 0.02987470105290413, + 0.03231706842780113, + 0.04267757385969162, + 0.029328521341085434, + 0.05465404689311981, + 0.03191159665584564, + 0.013437002897262573, + 0.010619699023663998, + 0.009746449068188667, + 0.012099078856408596, + 0.009265989065170288, + 0.01671484299004078, + 0.01145231258124113, + 0.011101345531642437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.03190779313445091 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295854300260544 + ], + "opposite_logprob_tokens": [ + -2.2379584312438965 + ], + "logprob_reward": -0.11295854300260544, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295854300260544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05144841969013214, + 0.14856885373592377, + 0.060680143535137177, + 0.04404854029417038, + 0.07199297100305557, + 0.19257137179374695, + 0.06145438924431801, + 0.1730293333530426, + 0.09039364010095596, + 0.10581232607364655 + ], + "credit_map": [ + 0.004806088283658028, + 0.0073645529337227345, + 0.004453336354345083, + 0.008548934943974018, + 0.007111845072358847, + 0.0064941635355353355, + 0.008600820787250996, + 0.0040686787106096745, + 0.011631613597273827, + 0.022574855014681816, + 0.013552109710872173, + 0.01803017035126686, + 0.01907222345471382, + 0.027423277497291565, + 0.024201104417443275, + 0.012083487585186958, + 0.006038981024175882, + 0.0073275272734463215, + 0.011064382269978523, + 0.009942927397787571, + 0.004878575447946787, + 0.0028176025953143835, + 0.012930329889059067, + 0.005679813679307699, + 0.004117609467357397, + 0.006189621053636074, + 0.004632233642041683, + 0.007091098465025425, + 0.0059736850671470165, + 0.0036274255253374577, + 0.009016765281558037, + 0.0034001015592366457, + 0.007999132387340069, + 0.007777599152177572, + 0.01199757494032383, + 0.015187286771833897, + 0.005821752827614546, + 0.004298363346606493, + 0.012536034919321537, + 0.006375228054821491, + 0.018463514745235443, + 0.023220742121338844, + 0.022635355591773987, + 0.023021690547466278, + 0.012344364076852798, + 0.03351517394185066, + 0.014540703035891056, + 0.044829826802015305, + 0.00553533248603344, + 0.00835656002163887, + 0.009716379456222057, + 0.00784536823630333, + 0.009550490416586399, + 0.006647179368883371, + 0.007951307110488415, + 0.005851766560226679, + 0.021396225318312645, + 0.01811278611421585, + 0.018347056582570076, + 0.02659059874713421, + 0.021241575479507446, + 0.017549429088830948, + 0.027266772463917732, + 0.022524874657392502, + 0.009404418058693409, + 0.01071175467222929, + 0.016031552106142044, + 0.013430910184979439, + 0.007282734848558903, + 0.011218735948204994, + 0.012235287576913834, + 0.010078240185976028, + 0.008435362949967384, + 0.010853360407054424, + 0.012936588376760483, + 0.013156708329916, + 0.010165448300540447, + 0.021377956494688988, + 0.013857212848961353, + 0.015029684640467167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.11295854300260544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202351681888103 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014202351681888103, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202351681888103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0771637111902237, + 0.16918474435806274, + 0.04176393151283264, + 0.09442388266324997, + 0.1511288583278656, + 0.1124524399638176, + 0.08378041535615921, + 0.08936751633882523, + 0.09776078909635544, + 0.08297371864318848 + ], + "credit_map": [ + 0.006600383203476667, + 0.01107875071465969, + 0.00825021043419838, + 0.012697343714535236, + 0.007632613182067871, + 0.00841042585670948, + 0.01583024486899376, + 0.006663731765002012, + 0.012845584191381931, + 0.018018430098891258, + 0.022842304781079292, + 0.02923029474914074, + 0.015935659408569336, + 0.027873652055859566, + 0.02728477492928505, + 0.015154043212532997, + 0.00294087128713727, + 0.005106409080326557, + 0.006601603236049414, + 0.006821092218160629, + 0.004461204167455435, + 0.0023237874265760183, + 0.009307652711868286, + 0.004201310221105814, + 0.007927046157419682, + 0.01716277003288269, + 0.009965965524315834, + 0.013084158301353455, + 0.018207788467407227, + 0.007538940757513046, + 0.012492057867348194, + 0.008045154623687267, + 0.01141874399036169, + 0.01233869232237339, + 0.019528884440660477, + 0.042937152087688446, + 0.007977576926350594, + 0.009224751964211464, + 0.03519335761666298, + 0.012509694322943687, + 0.013103579171001911, + 0.016627827659249306, + 0.011647080071270466, + 0.015941843390464783, + 0.010985353961586952, + 0.013881990686058998, + 0.011580894701182842, + 0.018683867529034615, + 0.008540570735931396, + 0.010074898600578308, + 0.012998813763260841, + 0.008368276059627533, + 0.016031822189688683, + 0.009801973588764668, + 0.008652834221720695, + 0.009311225265264511, + 0.01626906730234623, + 0.009057026356458664, + 0.01232190988957882, + 0.01118782814592123, + 0.007727290503680706, + 0.00881070364266634, + 0.01188820879906416, + 0.012105478905141354, + 0.008335061371326447, + 0.006331063341349363, + 0.013622024096548557, + 0.010891404934227467, + 0.014381006360054016, + 0.010062743909657001, + 0.021852215752005577, + 0.012285275384783745, + 0.006359109655022621, + 0.007524681743234396, + 0.017230791971087456, + 0.012926308438181877, + 0.008561096154153347, + 0.011586116626858711, + 0.007899832911789417, + 0.010885785333812237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.014202351681888103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031905481591820717 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.0031905481591820717, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031905481591820717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0549662709236145, + 0.04869266226887703, + 0.044764503836631775, + 0.04593485966324806, + 0.05934828892350197, + 0.32997870445251465, + 0.02441544085741043, + 0.029961753636598587, + 0.2785494327545166, + 0.0833880752325058 + ], + "credit_map": [ + 0.005132691469043493, + 0.007225408684462309, + 0.0060441214591264725, + 0.009018516167998314, + 0.006121722050011158, + 0.005261905491352081, + 0.011845959350466728, + 0.004315949510782957, + 0.0032196855172514915, + 0.005004757083952427, + 0.0067487140186131, + 0.007639991119503975, + 0.00535399466753006, + 0.008160335011780262, + 0.007939955219626427, + 0.004625224974006414, + 0.00241279904730618, + 0.004151345230638981, + 0.012102610431611538, + 0.00688527338206768, + 0.00326928892172873, + 0.002473063301295042, + 0.007439266890287399, + 0.00603085570037365, + 0.003693194128572941, + 0.0059657772071659565, + 0.006577547639608383, + 0.007605887949466705, + 0.007519397884607315, + 0.0031257851514965296, + 0.007262228988111019, + 0.004185040015727282, + 0.007413934450596571, + 0.006539039313793182, + 0.009806574322283268, + 0.011905006133019924, + 0.004316987935453653, + 0.004070352762937546, + 0.010737551376223564, + 0.004558844491839409, + 0.03434240445494652, + 0.04125941917300224, + 0.03371038660407066, + 0.04288618266582489, + 0.02167324721813202, + 0.05373930186033249, + 0.031132910400629044, + 0.0712348222732544, + 0.002669327659532428, + 0.0035449948627501726, + 0.0033530332148075104, + 0.002463280688971281, + 0.004067424684762955, + 0.002499152673408389, + 0.002900504507124424, + 0.0029177204705774784, + 0.003487006528303027, + 0.00337930372916162, + 0.004456609953194857, + 0.003999064210802317, + 0.004227120894938707, + 0.0033194858115166426, + 0.0036036986857652664, + 0.003489465918391943, + 0.02083955705165863, + 0.018426546826958656, + 0.02621752768754959, + 0.0278095081448555, + 0.0560426265001297, + 0.019078336656093597, + 0.08090449124574661, + 0.02923085354268551, + 0.008749153465032578, + 0.007902177050709724, + 0.01411798782646656, + 0.010075000114738941, + 0.006378142163157463, + 0.011474726721644402, + 0.0100124292075634, + 0.014678454026579857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.0031905481591820717 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera zooming in toward the rear tire of a Lamborghini Aventador SVJ?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car matte black carbon fiber with sharp lines and exaggerated aerodynamics?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the rear tire spinning rapidly while emitting thick white smoke and rubber particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rainy cyberpunk metropolis with neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944667384028435 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944667384028435, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944667384028435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16486378014087677, + 0.12778182327747345, + 0.03004402667284012, + 0.027643607929348946, + 0.05604119226336479, + 0.16999466717243195, + 0.04264000430703163, + 0.17839910089969635, + 0.13627465069293976, + 0.0663171112537384 + ], + "credit_map": [ + 0.011730003170669079, + 0.023418720811605453, + 0.01585637405514717, + 0.02751564048230648, + 0.015927305445075035, + 0.02425377257168293, + 0.03041389398276806, + 0.015748057514429092, + 0.008845268748700619, + 0.013857845216989517, + 0.010915882885456085, + 0.021868526935577393, + 0.015168800950050354, + 0.024690616875886917, + 0.021628955379128456, + 0.01080592256039381, + 0.002363926265388727, + 0.0031451962422579527, + 0.004461211152374744, + 0.005441547371447086, + 0.0034557764884084463, + 0.0016715624369680882, + 0.006876850500702858, + 0.0026279576122760773, + 0.003136369865387678, + 0.005588891915977001, + 0.002954392693936825, + 0.004424399230629206, + 0.0023560139816254377, + 0.0018896067049354315, + 0.004500439390540123, + 0.002793492516502738, + 0.005986435804516077, + 0.011382005177438259, + 0.00709872180595994, + 0.008093730546534061, + 0.004889375530183315, + 0.003948862198740244, + 0.009055906906723976, + 0.005586157087236643, + 0.013450810685753822, + 0.01810232736170292, + 0.02340400032699108, + 0.029069507494568825, + 0.008143973536789417, + 0.021425234153866768, + 0.020146772265434265, + 0.03625203296542168, + 0.007511472329497337, + 0.005042686127126217, + 0.0059818378649652, + 0.007188862655311823, + 0.003942703362554312, + 0.004695199429988861, + 0.004439088050276041, + 0.003838152624666691, + 0.02002626657485962, + 0.021269541233778, + 0.022146020084619522, + 0.026619670912623405, + 0.019794030115008354, + 0.018729185685515404, + 0.024583209306001663, + 0.02523118071258068, + 0.008173559792339802, + 0.011182702146470547, + 0.010466496460139751, + 0.014126460067927837, + 0.030588077381253242, + 0.012746837921440601, + 0.035440269857645035, + 0.013550255447626114, + 0.007287378888577223, + 0.005860502831637859, + 0.01006262470036745, + 0.012203441932797432, + 0.008605142123997211, + 0.008345982059836388, + 0.00840836949646473, + 0.00554366922006011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.005944667384028435 + } + ], + "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/0000131/group_6/index_tensor([4870, 4870])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_tensor([4870, 4870])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1637d3033967dca82d61d6959bde3b8815111a00 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_tensor([4870, 4870])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, high-angle, static shot focuses on a quiet, exhausted scene. A figure's head rests heavily on the surface of a table, cheek pressed against the cool, polished, dark wood. Long, dark hair, slightly disheveled, fans out across the intricate, swirling wood grain, its texture accentuated by the warm, low-angle light from a nearby lamp. In the foreground, a single, scattered white paper lies flat on the table. The core action is deconstructed into a subtle, rhythmic sequence: with each deep, steady exhalation of the unseen figure, a gentle puff of air moves across the smooth surface of the table, causing the edges of the paper to gently lift and flutter for a moment before settling back down in the quiet intervals between breaths. The scene is dominated by a warm palette of rich browns and soft golds, creating a strong play of light and shadow that enhances the sense of intimacy and solitude. The overall atmosphere is one of profound tranquility and deep contemplation, the only movement being the gentle, repetitive stirring of the paper, a visual anchor for the slow passage of time.", + "seed": 17361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_4870_seed_17361_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": -0.06416074368171394, + "vlm_logprob": -0.06416074368171394, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.06416074368171394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027401771396398544 + ], + "opposite_logprob_tokens": [ + -3.7774016857147217 + ], + "logprob_reward": -0.027401771396398544, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027401771396398544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44558459520339966, + 0.1235077828168869, + 0.07804398983716965, + 0.05615187808871269, + 0.037558574229478836, + 0.03897244483232498, + 0.028590397909283638, + 0.0386941023170948, + 0.04923079162836075, + 0.10366543382406235 + ], + "credit_map": [ + 0.05105173960328102, + 0.06810948997735977, + 0.035248395055532455, + 0.06674361974000931, + 0.06273919343948364, + 0.0575689971446991, + 0.057184092700481415, + 0.04693906754255295, + 0.012894703075289726, + 0.012466990388929844, + 0.007699429523199797, + 0.017287079244852066, + 0.01655689813196659, + 0.017095183953642845, + 0.023891262710094452, + 0.01561623066663742, + 0.007266351953148842, + 0.007006434723734856, + 0.005148926749825478, + 0.010706319473683834, + 0.009556780569255352, + 0.011280524544417858, + 0.015729274600744247, + 0.01134937722235918, + 0.00559656647965312, + 0.004007609095424414, + 0.003390508471056819, + 0.009021122939884663, + 0.006926003843545914, + 0.008120926097035408, + 0.01286437176167965, + 0.0062247677706182, + 0.003969346638768911, + 0.002688026987016201, + 0.0028386281337589025, + 0.005577195901423693, + 0.004569157026708126, + 0.005522154737263918, + 0.007880565710365772, + 0.0045134988613426685, + 0.004550422541797161, + 0.0031023721676319838, + 0.003988499287515879, + 0.008346282877027988, + 0.004344248212873936, + 0.005191172007471323, + 0.004295521881431341, + 0.005153927952051163, + 0.0028452537953853607, + 0.002921913517639041, + 0.002107439562678337, + 0.004405397456139326, + 0.003332641441375017, + 0.004002190660685301, + 0.005825393367558718, + 0.0031501685734838247, + 0.004398639313876629, + 0.0032866932451725006, + 0.0022049294784665108, + 0.005580288823693991, + 0.004554763901978731, + 0.005800149869173765, + 0.007617685943841934, + 0.005250950343906879, + 0.005582996178418398, + 0.004065158776938915, + 0.0034685104619711637, + 0.010267947800457478, + 0.005313543602824211, + 0.006993697490543127, + 0.007475742604583502, + 0.006063198205083609, + 0.008206749334931374, + 0.006098798010498285, + 0.008566231466829777, + 0.023953275755047798, + 0.011260353960096836, + 0.015719639137387276, + 0.011344530619680882, + 0.018515856936573982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.027401771396398544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029829446226358414 + ], + "opposite_logprob_tokens": [ + -3.529829502105713 + ], + "logprob_reward": -0.029829446226358414, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029829446226358414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36895182728767395, + 0.13826942443847656, + 0.05719343572854996, + 0.027131492272019386, + 0.052451927214860916, + 0.04465635120868683, + 0.12435755878686905, + 0.09816902130842209, + 0.029160097241401672, + 0.059658970683813095 + ], + "credit_map": [ + 0.049522608518600464, + 0.03364783897995949, + 0.03493840619921684, + 0.07654888927936554, + 0.037473950535058975, + 0.05082686245441437, + 0.041049595922231674, + 0.044943664222955704, + 0.014160136692225933, + 0.02186911553144455, + 0.011465433053672314, + 0.017626479268074036, + 0.018862735480070114, + 0.01815764047205448, + 0.020460067316889763, + 0.01566781848669052, + 0.00792571622878313, + 0.006277171894907951, + 0.005131969694048166, + 0.006253722123801708, + 0.00828490499407053, + 0.008127463050186634, + 0.00708518223837018, + 0.008107301779091358, + 0.003637510584667325, + 0.002428692067041993, + 0.00231871847063303, + 0.003015343099832535, + 0.0037498490419238806, + 0.0038189690094441175, + 0.004570980090647936, + 0.003591430839151144, + 0.005780333653092384, + 0.007353209890425205, + 0.00400801794603467, + 0.0077529363334178925, + 0.006209054030478001, + 0.007429251912981272, + 0.006886643823236227, + 0.007032477296888828, + 0.004406875465065241, + 0.003772617084905505, + 0.003969003912061453, + 0.008287666365504265, + 0.004918122664093971, + 0.006052244454622269, + 0.006415388081222773, + 0.006834437604993582, + 0.014866401441395283, + 0.017046842724084854, + 0.009712863713502884, + 0.0147575493901968, + 0.014068174175918102, + 0.017806682735681534, + 0.016182057559490204, + 0.01991697959601879, + 0.00943020824342966, + 0.012377073056995869, + 0.007927138358354568, + 0.016224145889282227, + 0.010882142931222916, + 0.013053825125098228, + 0.0140262171626091, + 0.014248274266719818, + 0.0027683707885444164, + 0.0031287965830415487, + 0.001731164869852364, + 0.004580354783684015, + 0.0033394431229680777, + 0.004217180423438549, + 0.00487342057749629, + 0.004521363880485296, + 0.00628426019102335, + 0.004993435461074114, + 0.005182615946978331, + 0.00852951779961586, + 0.006770493928343058, + 0.009203609079122543, + 0.008400491438806057, + 0.010294544510543346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.029829446226358414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646904490888119 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008646904490888119, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646904490888119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12463725358247757, + 0.1094868928194046, + 0.08086660504341125, + 0.051642417907714844, + 0.10849761217832565, + 0.1599065065383911, + 0.01441757008433342, + 0.031142007559537888, + 0.033514585345983505, + 0.28588855266571045 + ], + "credit_map": [ + 0.013609686866402626, + 0.01344764232635498, + 0.013677646405994892, + 0.022644365206360817, + 0.011506670154631138, + 0.01655232533812523, + 0.018668392673134804, + 0.014530523680150509, + 0.011796865612268448, + 0.010453973896801472, + 0.0183469969779253, + 0.020518053323030472, + 0.008436162024736404, + 0.01446477696299553, + 0.010703383013606071, + 0.014766685664653778, + 0.008778799325227737, + 0.008188704028725624, + 0.014019611291587353, + 0.01424408983439207, + 0.006629742681980133, + 0.010213788598775864, + 0.008227848447859287, + 0.010564016178250313, + 0.0038799901958554983, + 0.0033597787842154503, + 0.0044436464086174965, + 0.013076952658593655, + 0.004571553785353899, + 0.007035338319838047, + 0.006450170185416937, + 0.008824987336993217, + 0.011898760683834553, + 0.0136648528277874, + 0.01687479205429554, + 0.018286405131220818, + 0.010721908882260323, + 0.013856128789484501, + 0.009038771502673626, + 0.014155994169414043, + 0.017561953514814377, + 0.014896958135068417, + 0.020987993106245995, + 0.03071618638932705, + 0.013993151485919952, + 0.02234721928834915, + 0.012561172246932983, + 0.026841875165700912, + 0.001385191921144724, + 0.0015807641902938485, + 0.0011910040630027652, + 0.002453072927892208, + 0.0015856827376410365, + 0.0019839629530906677, + 0.002359329489991069, + 0.0018785622669383883, + 0.002496141940355301, + 0.002239000517874956, + 0.002629985101521015, + 0.005704652518033981, + 0.002831247402355075, + 0.004793289117515087, + 0.004140219185501337, + 0.006307472940534353, + 0.002621575491502881, + 0.002358976984396577, + 0.002390781184658408, + 0.006522563751786947, + 0.0031682997941970825, + 0.005184105131775141, + 0.004908272530883551, + 0.006360010709613562, + 0.021934160962700844, + 0.017465487122535706, + 0.031903889030218124, + 0.06219981610774994, + 0.023714076727628708, + 0.04488936439156532, + 0.02699090726673603, + 0.056790828704833984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.008646904490888119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521069347858429 + ], + "opposite_logprob_tokens": [ + -1.5021069049835205 + ], + "logprob_reward": -0.2521069347858429, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521069347858429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2344515323638916, + 0.06562410295009613, + 0.04997715726494789, + 0.04613287001848221, + 0.039026856422424316, + 0.06036994978785515, + 0.07776086777448654, + 0.18069575726985931, + 0.14487870037555695, + 0.1010822281241417 + ], + "credit_map": [ + 0.03241794928908348, + 0.018385302275419235, + 0.01579136773943901, + 0.050957780331373215, + 0.02908392623066902, + 0.03191297873854637, + 0.02587049826979637, + 0.030031709000468254, + 0.007432263810187578, + 0.005435981787741184, + 0.006889459677040577, + 0.008965244516730309, + 0.006371987983584404, + 0.009503239765763283, + 0.01086274441331625, + 0.01016317494213581, + 0.005106645170599222, + 0.005346275866031647, + 0.004108752124011517, + 0.006098074838519096, + 0.006067533046007156, + 0.007156165316700935, + 0.0089183384552598, + 0.007175367325544357, + 0.005318312905728817, + 0.006001072470098734, + 0.004831045400351286, + 0.007404393516480923, + 0.004996706731617451, + 0.006509341299533844, + 0.004770942963659763, + 0.006301052402704954, + 0.0042984071187675, + 0.004585874732583761, + 0.0028403284959495068, + 0.005420069210231304, + 0.00580283347517252, + 0.005174774676561356, + 0.005685123614966869, + 0.0052194418385624886, + 0.006811860017478466, + 0.005129140801727772, + 0.0061950585804879665, + 0.011281690560281277, + 0.005255293101072311, + 0.008875715546309948, + 0.006917732302099466, + 0.009903453290462494, + 0.0077869282104074955, + 0.011886529624462128, + 0.006280233617872, + 0.01581895351409912, + 0.007893050089478493, + 0.011209835298359394, + 0.006953663658350706, + 0.009931663982570171, + 0.017497817054390907, + 0.022204170003533363, + 0.014817886054515839, + 0.028822582215070724, + 0.018801838159561157, + 0.026591721922159195, + 0.02149948850274086, + 0.030460244044661522, + 0.01309218443930149, + 0.016030892729759216, + 0.011691998690366745, + 0.026473481208086014, + 0.014568996615707874, + 0.020732512697577477, + 0.01795876957476139, + 0.024329854175448418, + 0.00886451918631792, + 0.006470838095992804, + 0.007215370424091816, + 0.021435391157865524, + 0.008956929668784142, + 0.016270358115434647, + 0.012209408916532993, + 0.01965940371155739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.2521069347858429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.4201519191265106, + 0.1150396317243576, + 0.08656974881887436, + 0.053439658135175705, + 0.05440256744623184, + 0.04966649413108826, + 0.04298748821020126, + 0.025797365233302116, + 0.0425407811999321, + 0.10940434038639069 + ], + "credit_map": [ + 0.05371108651161194, + 0.057568736374378204, + 0.03180164098739624, + 0.060442354530096054, + 0.05497782304883003, + 0.054650578647851944, + 0.05681958794593811, + 0.05018014460802078, + 0.01204106118530035, + 0.014146717265248299, + 0.011876771226525307, + 0.018178531900048256, + 0.014225534163415432, + 0.014657982625067234, + 0.015968790277838707, + 0.013944253325462341, + 0.009124070405960083, + 0.009451503865420818, + 0.009272372350096703, + 0.011777537874877453, + 0.01214591134339571, + 0.01079775020480156, + 0.013033955357968807, + 0.01096665021032095, + 0.005486241076141596, + 0.005946691147983074, + 0.00440696906298399, + 0.007953965105116367, + 0.007078475318849087, + 0.0066251554526388645, + 0.00949829537421465, + 0.006443871185183525, + 0.005092382896691561, + 0.007485284004360437, + 0.004543522838503122, + 0.008367885835468769, + 0.0074073635041713715, + 0.007588433101773262, + 0.006776159163564444, + 0.007141538895666599, + 0.005650022532790899, + 0.005671538878232241, + 0.00320077920332551, + 0.007808192167431116, + 0.005489261820912361, + 0.00746593764051795, + 0.007808045484125614, + 0.00657272431999445, + 0.00555466627702117, + 0.00474819540977478, + 0.0030576454009860754, + 0.0069315796718001366, + 0.0056223031133413315, + 0.00576352421194315, + 0.006115906406193972, + 0.00519367353990674, + 0.002987977582961321, + 0.003732910845428705, + 0.002496895147487521, + 0.003617923939600587, + 0.0031373361125588417, + 0.003535150783136487, + 0.002998329931870103, + 0.003290843451395631, + 0.0038272731471806765, + 0.004258453845977783, + 0.003091474762186408, + 0.00936164427548647, + 0.003935112152248621, + 0.0058615850284695625, + 0.00586014986038208, + 0.0063450876623392105, + 0.010933131910860538, + 0.0093936612829566, + 0.0059429616667330265, + 0.0173103678971529, + 0.011949117295444012, + 0.01638173870742321, + 0.024013271555304527, + 0.013480102643370628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0028186615090817213 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle, static shot focuses on a quiet, exhausted scene. A figure's head rests heavily on the surface of a table, cheek pressed against the cool, polished, dark wood. Long, dark hair, slightly disheveled, fans out across the intricate, swirling wood grain, its texture accentuated by the warm, low-angle light from a nearby lamp. In the foreground, a single, scattered white paper lies flat on the table. The core action is deconstructed into a subtle, rhythmic sequence: with each deep, steady exhalation of the unseen figure, a gentle puff of air moves across the smooth surface of the table, causing the edges of the paper to gently lift and flutter for a moment before settling back down in the quiet intervals between breaths. The scene is dominated by a warm palette of rich browns and soft golds, creating a strong play of light and shadow that enhances the sense of intimacy and solitude. The overall atmosphere is one of profound tranquility and deep contemplation, the only movement being the gentle, repetitive stirring of the paper, a visual anchor for the slow passage of time.", + "seed": 17362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_4870_seed_17362_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": -0.013243385404348374, + "vlm_logprob": -0.013243385404348374, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.013243385404348374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02690371498465538 + ], + "opposite_logprob_tokens": [ + -3.7769036293029785 + ], + "logprob_reward": -0.02690371498465538, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02690371498465538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2854098081588745, + 0.049294713884592056, + 0.026496972888708115, + 0.11369916051626205, + 0.01860961690545082, + 0.023901863023638725, + 0.0980854406952858, + 0.017534729093313217, + 0.12147343903779984, + 0.24549423158168793 + ], + "credit_map": [ + 0.029455263167619705, + 0.021678414195775986, + 0.0477718748152256, + 0.018826071172952652, + 0.08330609649419785, + 0.03451588749885559, + 0.017107823863625526, + 0.03274838253855705, + 0.003455085912719369, + 0.0038412739522755146, + 0.004320578649640083, + 0.002382055390626192, + 0.008882669731974602, + 0.013844408094882965, + 0.004290384706109762, + 0.008278257213532925, + 0.0028740763664245605, + 0.0030712706502527, + 0.0016764529282227159, + 0.0020530023612082005, + 0.0037340077105909586, + 0.005504664499312639, + 0.003031972562894225, + 0.004551525693386793, + 0.011020056903362274, + 0.007353045511990786, + 0.014442004263401031, + 0.0062262085266411304, + 0.035686179995536804, + 0.016511790454387665, + 0.007980160415172577, + 0.014479725621640682, + 0.0016460672486573458, + 0.0013334216782823205, + 0.0018326598219573498, + 0.0010912176221609116, + 0.004582679830491543, + 0.003910538740456104, + 0.0015009157359600067, + 0.0027121163439005613, + 0.0020341547206044197, + 0.001884255325421691, + 0.003289740066975355, + 0.00222078082151711, + 0.005987060721963644, + 0.002998341340571642, + 0.002708917483687401, + 0.002778611844405532, + 0.007852002047002316, + 0.0056152050383389, + 0.0154531579464674, + 0.0036499069537967443, + 0.03353681415319443, + 0.01156249362975359, + 0.007344300393015146, + 0.013071565888822079, + 0.0015132615808397532, + 0.0017466722056269646, + 0.0018087461357936263, + 0.0015110453823581338, + 0.003124001668766141, + 0.0037378377746790648, + 0.001818617805838585, + 0.002274546306580305, + 0.009568345732986927, + 0.01836356148123741, + 0.006812288425862789, + 0.008189614862203598, + 0.022087031975388527, + 0.031313035637140274, + 0.014468193054199219, + 0.01067136786878109, + 0.006195503752678633, + 0.014469515532255173, + 0.004657142795622349, + 0.004497903864830732, + 0.022958967834711075, + 0.10662348568439484, + 0.00904349610209465, + 0.07704819738864899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02690371498465538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210226014256477 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210226014256477, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210226014256477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06429412961006165, + 0.058096952736377716, + 0.07376362383365631, + 0.16696563363075256, + 0.1068386361002922, + 0.10446929186582565, + 0.05122197046875954, + 0.06091698631644249, + 0.03672182932496071, + 0.2767109274864197 + ], + "credit_map": [ + 0.00788028258830309, + 0.007083038333803415, + 0.0050852638669312, + 0.006549423560500145, + 0.008374081924557686, + 0.010785176418721676, + 0.008690119720995426, + 0.009846745990216732, + 0.006446169689297676, + 0.00576808862388134, + 0.0076788607984781265, + 0.004309084266424179, + 0.010168347507715225, + 0.011136699467897415, + 0.005796707235276699, + 0.0067929974757134914, + 0.01000170223414898, + 0.0075974357314407825, + 0.0062844157218933105, + 0.006560306064784527, + 0.011819896288216114, + 0.013299146667122841, + 0.009237943217158318, + 0.008962778374552727, + 0.022192807868123055, + 0.018212754279375076, + 0.014907910488545895, + 0.015221924521028996, + 0.0297965407371521, + 0.027514370158314705, + 0.017738858237862587, + 0.0213804692029953, + 0.010896972380578518, + 0.010655505582690239, + 0.009576325304806232, + 0.00808370765298605, + 0.024632528424263, + 0.019233303144574165, + 0.010456731542944908, + 0.013303558342158794, + 0.011306402273476124, + 0.013973464258015156, + 0.010132329538464546, + 0.007246107794344425, + 0.014152053743600845, + 0.0192712415009737, + 0.009443456307053566, + 0.018944231793284416, + 0.0038549057208001614, + 0.0028624918777495623, + 0.007287804502993822, + 0.0016537162009626627, + 0.01776190660893917, + 0.00603999150916934, + 0.003512631170451641, + 0.008248521015048027, + 0.0053126150742173195, + 0.0076073287054896355, + 0.005934784654527903, + 0.0038679966237396, + 0.008411005139350891, + 0.008074924349784851, + 0.008311371318995953, + 0.013396958820521832, + 0.0038319912273436785, + 0.007187637500464916, + 0.002800452522933483, + 0.004195862449705601, + 0.0038639253471046686, + 0.006603811867535114, + 0.0044975318014621735, + 0.0037406147457659245, + 0.02521597221493721, + 0.02663516439497471, + 0.01210842840373516, + 0.014090961776673794, + 0.04068022966384888, + 0.06646642088890076, + 0.024666959419846535, + 0.06684678047895432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.014210226014256477 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02614464983344078, + 0.027011286467313766, + 0.02476346492767334, + 0.03930317237973213, + 0.016099590808153152, + 0.016623131930828094, + 0.3645223379135132, + 0.08151363581418991, + 0.0758322924375534, + 0.3281864523887634 + ], + "credit_map": [ + 0.0033629718236625195, + 0.002211199142038822, + 0.0019454921130090952, + 0.002127087442204356, + 0.0044088782742619514, + 0.004660885781049728, + 0.0026437186170369387, + 0.0047844150103628635, + 0.0028844636399298906, + 0.0026197500992566347, + 0.0030138101428747177, + 0.0022342572920024395, + 0.005664142780005932, + 0.0036786592099815607, + 0.003805088112130761, + 0.00311111519113183, + 0.002288896357640624, + 0.0017441571690142155, + 0.0027220556512475014, + 0.001639888621866703, + 0.005742935463786125, + 0.0036862969864159822, + 0.003634669119492173, + 0.003304565791040659, + 0.0033292381558567286, + 0.0021426777821034193, + 0.006253927014768124, + 0.0021220624912530184, + 0.011982958763837814, + 0.005986761301755905, + 0.002759933937340975, + 0.004725610837340355, + 0.0016223413404077291, + 0.0017908102599903941, + 0.0011061810655519366, + 0.0014991973293945193, + 0.0030114788096398115, + 0.0028156356420367956, + 0.001869985950179398, + 0.0023839599452912807, + 0.001666353433392942, + 0.0013254202203825116, + 0.002535391366109252, + 0.0008111366769298911, + 0.0038064452819526196, + 0.002416136907413602, + 0.0015087261563166976, + 0.0025535232853144407, + 0.030400581657886505, + 0.019574221223592758, + 0.06196464225649834, + 0.012448981404304504, + 0.12290999293327332, + 0.04382863640785217, + 0.026038117706775665, + 0.047357186675071716, + 0.005663644056767225, + 0.008171509020030499, + 0.004283458925783634, + 0.005132139194756746, + 0.016632307320833206, + 0.0131455073133111, + 0.012556931935250759, + 0.015928134322166443, + 0.008421383798122406, + 0.009896047413349152, + 0.005636394023895264, + 0.008702518418431282, + 0.009374293498694897, + 0.014705304056406021, + 0.010982926934957504, + 0.008113419637084007, + 0.011701441369950771, + 0.022098492830991745, + 0.00880367960780859, + 0.009063259698450565, + 0.024398263543844223, + 0.1233178973197937, + 0.022089822217822075, + 0.10671360790729523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.004097756464034319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188826739788055 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.018188826739788055, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188826739788055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08913746476173401, + 0.04723586514592171, + 0.032658111304044724, + 0.03130055218935013, + 0.04182631894946098, + 0.02780751883983612, + 0.186138316988945, + 0.061488308012485504, + 0.07246582955121994, + 0.40994176268577576 + ], + "credit_map": [ + 0.007444872986525297, + 0.00831377599388361, + 0.011404413729906082, + 0.006899747997522354, + 0.017097769305109978, + 0.017450673505663872, + 0.008290767669677734, + 0.012235438451170921, + 0.004166072700172663, + 0.004289987962692976, + 0.005251020193099976, + 0.003009561449289322, + 0.008721736259758472, + 0.010115640237927437, + 0.005655067972838879, + 0.006026777438819408, + 0.003851167857646942, + 0.0024168239906430244, + 0.0027127449866384268, + 0.0015368204331025481, + 0.006122200284153223, + 0.006474792957305908, + 0.0035807734820991755, + 0.005962786264717579, + 0.0042001036927104, + 0.0028813788667321205, + 0.003588903695344925, + 0.002066770801320672, + 0.007571677211672068, + 0.004118364304304123, + 0.003121068701148033, + 0.0037522867787629366, + 0.004301808774471283, + 0.004327105823904276, + 0.004352128133177757, + 0.0023576959501951933, + 0.010356982238590717, + 0.0062442924827337265, + 0.0037106105592101812, + 0.006175692658871412, + 0.0024428407195955515, + 0.0020917002111673355, + 0.00292778667062521, + 0.002207797719165683, + 0.005597854498773813, + 0.00420643575489521, + 0.0038721782620996237, + 0.004460924305021763, + 0.015384864062070847, + 0.010119233280420303, + 0.03133909031748772, + 0.006470328196883202, + 0.06304436177015305, + 0.021749338135123253, + 0.013303954154253006, + 0.02472713217139244, + 0.003920668736100197, + 0.005717037711292505, + 0.0032181323040276766, + 0.00423494353890419, + 0.014619821682572365, + 0.009514490142464638, + 0.010137793608009815, + 0.010125421918928623, + 0.006686632987111807, + 0.00705665023997426, + 0.00586855411529541, + 0.00846001598984003, + 0.01010347530245781, + 0.013625221326947212, + 0.012170576490461826, + 0.00849470030516386, + 0.03796413168311119, + 0.04897097870707512, + 0.01782909967005253, + 0.01975335367023945, + 0.06521045416593552, + 0.11874181777238846, + 0.029259832575917244, + 0.0722120851278305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.018188826739788055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816402819007635 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816402819007635, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816402819007635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15387961268424988, + 0.042546346783638, + 0.0260222926735878, + 0.16298162937164307, + 0.023753155022859573, + 0.03042236901819706, + 0.12443581968545914, + 0.03664242476224899, + 0.02589958719909191, + 0.3734166920185089 + ], + "credit_map": [ + 0.014356671832501888, + 0.011207024566829205, + 0.023987723514437675, + 0.008105381391942501, + 0.04420879855751991, + 0.019514156505465508, + 0.012122110463678837, + 0.02037775330245495, + 0.003934734500944614, + 0.0035456742625683546, + 0.005497234407812357, + 0.002621843945235014, + 0.011352590285241604, + 0.0057119629345834255, + 0.004401164595037699, + 0.005481143947690725, + 0.00231915689073503, + 0.0025433709379285574, + 0.002753007924184203, + 0.0017208842327818274, + 0.005552890710532665, + 0.004588582552969456, + 0.0025687431916594505, + 0.003975657746195793, + 0.015629049390554428, + 0.009447415359318256, + 0.02425789088010788, + 0.008436054922640324, + 0.05269599333405495, + 0.022990766912698746, + 0.010015588253736496, + 0.019508888944983482, + 0.002934362506493926, + 0.0028513288125395775, + 0.0024099135771393776, + 0.0017949518514797091, + 0.005137169733643532, + 0.0031815844122320414, + 0.002377096563577652, + 0.003066748147830367, + 0.0030204325448721647, + 0.0027143724728375673, + 0.004247049335390329, + 0.0017175687244161963, + 0.006674126721918583, + 0.005126035772264004, + 0.002427744911983609, + 0.004495040979236364, + 0.010359739884734154, + 0.0067579918541014194, + 0.02076488919556141, + 0.004409793298691511, + 0.041792962700128555, + 0.015378855168819427, + 0.009048138745129108, + 0.015923451632261276, + 0.0031550233252346516, + 0.0035315121058374643, + 0.0025201065000146627, + 0.00243442808277905, + 0.008994046598672867, + 0.007252535317093134, + 0.005043724086135626, + 0.003711051307618618, + 0.0037132275756448507, + 0.0033914928790181875, + 0.0019168892176821828, + 0.002583456691354513, + 0.0032838343176990747, + 0.003910152241587639, + 0.004213237203657627, + 0.00288729858584702, + 0.015854474157094955, + 0.023996353149414062, + 0.008853132836520672, + 0.008251084946095943, + 0.03375985100865364, + 0.15097157657146454, + 0.020869620144367218, + 0.11086063832044601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.002816402819007635 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle, static shot focuses on a quiet, exhausted scene. A figure's head rests heavily on the surface of a table, cheek pressed against the cool, polished, dark wood. Long, dark hair, slightly disheveled, fans out across the intricate, swirling wood grain, its texture accentuated by the warm, low-angle light from a nearby lamp. In the foreground, a single, scattered white paper lies flat on the table. The core action is deconstructed into a subtle, rhythmic sequence: with each deep, steady exhalation of the unseen figure, a gentle puff of air moves across the smooth surface of the table, causing the edges of the paper to gently lift and flutter for a moment before settling back down in the quiet intervals between breaths. The scene is dominated by a warm palette of rich browns and soft golds, creating a strong play of light and shadow that enhances the sense of intimacy and solitude. The overall atmosphere is one of profound tranquility and deep contemplation, the only movement being the gentle, repetitive stirring of the paper, a visual anchor for the slow passage of time.", + "seed": 17363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_4870_seed_17363_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": -0.031092424830421805, + "vlm_logprob": -0.031092424830421805, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.031092424830421805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03389621153473854 + ], + "opposite_logprob_tokens": [ + -3.5338962078094482 + ], + "logprob_reward": -0.03389621153473854, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03389621153473854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14310285449028015, + 0.056170377880334854, + 0.3212463855743408, + 0.09245780855417252, + 0.05465410649776459, + 0.04641461744904518, + 0.09483292698860168, + 0.05345496907830238, + 0.07723496854305267, + 0.06043101102113724 + ], + "credit_map": [ + 0.014905473217368126, + 0.021346434950828552, + 0.025510890409350395, + 0.027809705585241318, + 0.019180161878466606, + 0.012560714036226273, + 0.01570967771112919, + 0.00607979716733098, + 0.006712771952152252, + 0.007307203020900488, + 0.00813654437661171, + 0.008284712210297585, + 0.007050896529108286, + 0.005301197990775108, + 0.008499013260006905, + 0.004878040868788958, + 0.03908691555261612, + 0.044504694640636444, + 0.056493621319532394, + 0.056704942137002945, + 0.056639134883880615, + 0.020355194807052612, + 0.028153905645012856, + 0.019307976588606834, + 0.01269390806555748, + 0.01081454660743475, + 0.01187584176659584, + 0.01405594777315855, + 0.016985764726996422, + 0.007597608026117086, + 0.010722723789513111, + 0.007711464073508978, + 0.009870937094092369, + 0.008038501255214214, + 0.006433091126382351, + 0.008666915819048882, + 0.006766279693692923, + 0.004602053668349981, + 0.005616762675344944, + 0.004659563768655062, + 0.009898778982460499, + 0.005406176205724478, + 0.0066188485361635685, + 0.00573935080319643, + 0.0035480582155287266, + 0.0021322511602193117, + 0.0046179876662790775, + 0.00845316331833601, + 0.010830224491655827, + 0.01346126664429903, + 0.011737702414393425, + 0.017765777185559273, + 0.008482184261083603, + 0.0083406250923872, + 0.009809233248233795, + 0.014405918307602406, + 0.012269927188754082, + 0.009787260554730892, + 0.00754518061876297, + 0.008259620517492294, + 0.005254772957414389, + 0.0026121216360479593, + 0.0039039619732648134, + 0.003822120139375329, + 0.011452929116785526, + 0.007484727539122105, + 0.010228843428194523, + 0.015540258958935738, + 0.010221726261079311, + 0.004425222985446453, + 0.006828204728662968, + 0.011053058318793774, + 0.007347674574702978, + 0.0070396894589066505, + 0.007523926440626383, + 0.010996714234352112, + 0.006946628447622061, + 0.00351175875402987, + 0.008967483416199684, + 0.008097137324512005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.03389621153473854 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372852876782417 + ], + "opposite_logprob_tokens": [ + -3.40872859954834 + ], + "logprob_reward": -0.03372852876782417, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372852876782417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2414277344942093, + 0.22514306008815765, + 0.10128788650035858, + 0.10528016835451126, + 0.056615859270095825, + 0.05885535106062889, + 0.033622998744249344, + 0.03880847990512848, + 0.03980043902993202, + 0.09915801137685776 + ], + "credit_map": [ + 0.025675658136606216, + 0.03517414629459381, + 0.038049858063459396, + 0.04183206334710121, + 0.036403052508831024, + 0.02199774980545044, + 0.02939239889383316, + 0.012902786023914814, + 0.025501329451799393, + 0.03346637263894081, + 0.03380727767944336, + 0.04329835623502731, + 0.040606025606393814, + 0.011783883906900883, + 0.019594457000494003, + 0.0170853603631258, + 0.013845358043909073, + 0.018770189955830574, + 0.014549758285284042, + 0.01027654018253088, + 0.012881944887340069, + 0.006851931568235159, + 0.012285067699849606, + 0.011827097274363041, + 0.015035178512334824, + 0.015607904642820358, + 0.013731608167290688, + 0.01852702535688877, + 0.013344818726181984, + 0.00778060220181942, + 0.01176974643021822, + 0.009483291767537594, + 0.009415612556040287, + 0.00757468631491065, + 0.006930441129952669, + 0.005135773215442896, + 0.011428308673202991, + 0.0050745559856295586, + 0.005326808895915747, + 0.005729669705033302, + 0.01540440134704113, + 0.007101035211235285, + 0.008384088054299355, + 0.007195647805929184, + 0.007467188406735659, + 0.004030033480376005, + 0.003584139747545123, + 0.00568881630897522, + 0.0058418381959199905, + 0.004897563252598047, + 0.004509421065449715, + 0.005535353906452656, + 0.0034804055467247963, + 0.002816202351823449, + 0.003266403218731284, + 0.0032758142333477736, + 0.005989366210997105, + 0.004489803221076727, + 0.005348126869648695, + 0.006630648858845234, + 0.004086664877831936, + 0.002028900198638439, + 0.003834300907328725, + 0.006400672253221273, + 0.0063718766905367374, + 0.00469411863014102, + 0.0056599765084683895, + 0.007144947070628405, + 0.005435814615339041, + 0.002880627056583762, + 0.003370231483131647, + 0.0042428444139659405, + 0.013001362793147564, + 0.011762234382331371, + 0.011739716865122318, + 0.019361762329936028, + 0.010535688139498234, + 0.0071532209403812885, + 0.012585536576807499, + 0.01301848515868187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03372852876782417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292383179068565 + ], + "opposite_logprob_tokens": [ + -3.773292303085327 + ], + "logprob_reward": -0.023292383179068565, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292383179068565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06744692474603653, + 0.12041384726762772, + 0.1956826150417328, + 0.06564240902662277, + 0.0718919187784195, + 0.038777075707912445, + 0.05192263796925545, + 0.03150929510593414, + 0.11306209117174149, + 0.24365118145942688 + ], + "credit_map": [ + 0.006245079450309277, + 0.015031103044748306, + 0.01093754731118679, + 0.008054820820689201, + 0.011829088442027569, + 0.003872299799695611, + 0.006313383113592863, + 0.005163602065294981, + 0.012348609045147896, + 0.02228817157447338, + 0.01709519699215889, + 0.016889452934265137, + 0.016481071710586548, + 0.009332340210676193, + 0.015310862101614475, + 0.010668138042092323, + 0.01762768253684044, + 0.04231548681855202, + 0.032544758170843124, + 0.03223590925335884, + 0.018111441284418106, + 0.012341047637164593, + 0.023230070248246193, + 0.017276206985116005, + 0.009206501767039299, + 0.007376809138804674, + 0.008948463015258312, + 0.006691809743642807, + 0.012314324267208576, + 0.004691963084042072, + 0.009238699451088905, + 0.007173837162554264, + 0.013045843690633774, + 0.00941433385014534, + 0.00987158715724945, + 0.012538083828985691, + 0.009073303081095219, + 0.005320949014276266, + 0.006888784933835268, + 0.005739028099924326, + 0.004621484782546759, + 0.005600215867161751, + 0.00527641735970974, + 0.004602830857038498, + 0.007753568235784769, + 0.003295161994174123, + 0.003244183724746108, + 0.004383208695799112, + 0.006513148546218872, + 0.007626230362802744, + 0.005947416182607412, + 0.007570412941277027, + 0.006932252086699009, + 0.004739491268992424, + 0.0050657051615417, + 0.0075279781594872475, + 0.006329171359539032, + 0.0043984162621200085, + 0.00441329600289464, + 0.004101739265024662, + 0.0030652335844933987, + 0.0015366500010713935, + 0.0028446894139051437, + 0.004820096306502819, + 0.01691209152340889, + 0.01049028616398573, + 0.01439687144011259, + 0.019246650859713554, + 0.017319437116384506, + 0.0071945576928555965, + 0.01109403371810913, + 0.016408152878284454, + 0.024465691298246384, + 0.026762928813695908, + 0.029680386185646057, + 0.04706602171063423, + 0.030545491725206375, + 0.013650501146912575, + 0.03664589300751686, + 0.03483424335718155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.023292383179068565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620557926595211 + ], + "opposite_logprob_tokens": [ + -2.812055826187134 + ], + "logprob_reward": -0.0620557926595211, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620557926595211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2193121761083603, + 0.0727124884724617, + 0.15512092411518097, + 0.17551785707473755, + 0.08830653131008148, + 0.08217058330774307, + 0.041730452328920364, + 0.03140846639871597, + 0.05011835694313049, + 0.08360221982002258 + ], + "credit_map": [ + 0.020623642951250076, + 0.05898432433605194, + 0.0443369559943676, + 0.02679445967078209, + 0.025444643571972847, + 0.013094170950353146, + 0.01644936203956604, + 0.013584606349468231, + 0.007180855143815279, + 0.009741644375026226, + 0.012633493170142174, + 0.014940556138753891, + 0.011530043557286263, + 0.0040893591940402985, + 0.007671281695365906, + 0.004925251938402653, + 0.020531103014945984, + 0.023358523845672607, + 0.029215017333626747, + 0.03359774500131607, + 0.013988979160785675, + 0.009681429713964462, + 0.01441098190844059, + 0.010337124578654766, + 0.022809071466326714, + 0.017669932916760445, + 0.022219542413949966, + 0.02256368100643158, + 0.035783033818006516, + 0.011099289171397686, + 0.02488480508327484, + 0.018488479778170586, + 0.02444114349782467, + 0.013107161968946457, + 0.011748316697776318, + 0.008051743730902672, + 0.010619524866342545, + 0.005067064426839352, + 0.007494056597352028, + 0.007777512073516846, + 0.017262281849980354, + 0.009842565283179283, + 0.011144299991428852, + 0.011227204464375973, + 0.012027889490127563, + 0.004969940520823002, + 0.007107947487384081, + 0.008588442578911781, + 0.008166860789060593, + 0.0050454335287213326, + 0.005761630367487669, + 0.006430343259125948, + 0.0052509428933262825, + 0.002052913187071681, + 0.003641383023932576, + 0.005380942486226559, + 0.004462058190256357, + 0.004370668437331915, + 0.004207131918519735, + 0.004257577005773783, + 0.0035862254444509745, + 0.002674040151759982, + 0.0036394232884049416, + 0.0042113387025892735, + 0.00882222130894661, + 0.007055417634546757, + 0.007032887078821659, + 0.006840817164629698, + 0.007113604340702295, + 0.0035141382832080126, + 0.004370891489088535, + 0.005368378013372421, + 0.009848393499851227, + 0.010569530539214611, + 0.010563002899289131, + 0.016485651955008507, + 0.009968623518943787, + 0.0057700215838849545, + 0.010755197145044804, + 0.009641796350479126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0620557926595211 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489208010956645 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489208010956645, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489208010956645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22162100672721863, + 0.1507335901260376, + 0.1063927486538887, + 0.1181475967168808, + 0.07536699622869492, + 0.09510872513055801, + 0.05055079609155655, + 0.05065661668777466, + 0.04866618290543556, + 0.08275578916072845 + ], + "credit_map": [ + 0.025864511728286743, + 0.036563482135534286, + 0.041448671370744705, + 0.03972238302230835, + 0.026483040302991867, + 0.01688002608716488, + 0.020393749698996544, + 0.014265155419707298, + 0.015422414988279343, + 0.02333892695605755, + 0.025075113400816917, + 0.026662364602088928, + 0.02868940867483616, + 0.008689702488481998, + 0.013167222961783409, + 0.009688430465757847, + 0.01344901043921709, + 0.015301503241062164, + 0.017748991027474403, + 0.021184492856264114, + 0.014210286550223827, + 0.005716864950954914, + 0.010993396863341331, + 0.007788196671754122, + 0.014804261736571789, + 0.014029313810169697, + 0.01586010493338108, + 0.01530748512595892, + 0.02117237076163292, + 0.008957866579294205, + 0.016299936920404434, + 0.011716247536242008, + 0.011270706541836262, + 0.010795711539685726, + 0.009615978226065636, + 0.01007597055286169, + 0.008986336179077625, + 0.006449511740356684, + 0.009677833877503872, + 0.008494953624904156, + 0.013227934017777443, + 0.0134359709918499, + 0.012603533454239368, + 0.01386247668415308, + 0.018853601068258286, + 0.007446555886417627, + 0.007502558175474405, + 0.008176096715033054, + 0.007816869765520096, + 0.008590735495090485, + 0.00651170639321208, + 0.006622598972171545, + 0.008032986894249916, + 0.00403956975787878, + 0.0042829010635614395, + 0.00465342728421092, + 0.007393702864646912, + 0.007120446301996708, + 0.006130389869213104, + 0.007708216551691294, + 0.005861575249582529, + 0.00401123333722353, + 0.00658545782789588, + 0.005845595151185989, + 0.009442079812288284, + 0.008479679003357887, + 0.007250804919749498, + 0.005945983342826366, + 0.006563956383615732, + 0.0036427397280931473, + 0.0033477009274065495, + 0.0039932383224368095, + 0.011786531656980515, + 0.01021597534418106, + 0.010276141576468945, + 0.014154803939163685, + 0.010176064446568489, + 0.004875587299466133, + 0.010662115179002285, + 0.010608566924929619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.002489208010956645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, high-angle, static shot focuses on a quiet, exhausted scene. A figure's head rests heavily on the surface of a table, cheek pressed against the cool, polished, dark wood. Long, dark hair, slightly disheveled, fans out across the intricate, swirling wood grain, its texture accentuated by the warm, low-angle light from a nearby lamp. In the foreground, a single, scattered white paper lies flat on the table. The core action is deconstructed into a subtle, rhythmic sequence: with each deep, steady exhalation of the unseen figure, a gentle puff of air moves across the smooth surface of the table, causing the edges of the paper to gently lift and flutter for a moment before settling back down in the quiet intervals between breaths. The scene is dominated by a warm palette of rich browns and soft golds, creating a strong play of light and shadow that enhances the sense of intimacy and solitude. The overall atmosphere is one of profound tranquility and deep contemplation, the only movement being the gentle, repetitive stirring of the paper, a visual anchor for the slow passage of time.", + "seed": 17364, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_6/index_4870_seed_17364_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": -0.030590294464491307, + "vlm_logprob": -0.030590294464491307, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.030590294464491307, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0056923432275652885 + ], + "opposite_logprob_tokens": [ + -6.880692481994629 + ], + "logprob_reward": -0.0056923432275652885, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0056923432275652885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10975638031959534, + 0.20729567110538483, + 0.06303216516971588, + 0.29474228620529175, + 0.02008967101573944, + 0.03400694206357002, + 0.01282751839607954, + 0.1489313393831253, + 0.04268287494778633, + 0.0666351318359375 + ], + "credit_map": [ + 0.0075435154139995575, + 0.005814902484416962, + 0.004571455996483564, + 0.03745424374938011, + 0.00661825155839324, + 0.03399074450135231, + 0.0048114764504134655, + 0.008951784111559391, + 0.014526520855724812, + 0.0069037796929478645, + 0.007763626519590616, + 0.08655881136655807, + 0.005909688770771027, + 0.05655284225940704, + 0.005425495095551014, + 0.02365492656826973, + 0.004664148669689894, + 0.002547529060393572, + 0.002239026129245758, + 0.0259480532258749, + 0.002065726090222597, + 0.016580307856202126, + 0.001773224794305861, + 0.007214151322841644, + 0.021113121882081032, + 0.010096625424921513, + 0.00966972392052412, + 0.12296520173549652, + 0.010027552954852581, + 0.07878036797046661, + 0.007994859479367733, + 0.034094829112291336, + 0.0021225973032414913, + 0.0019749419298022985, + 0.0019354226533323526, + 0.004553152713924646, + 0.002500133356079459, + 0.003224909771233797, + 0.001723835477605462, + 0.0020546766463667154, + 0.002525392221286893, + 0.0024759829975664616, + 0.002689702669158578, + 0.008638850413262844, + 0.0032731532119214535, + 0.005532058887183666, + 0.004699178505688906, + 0.004172624088823795, + 0.0013041440397500992, + 0.0011941634584218264, + 0.0016287227626889944, + 0.0017932194750756025, + 0.0018433354562148452, + 0.0013807977084070444, + 0.0018115269485861063, + 0.0018716083141043782, + 0.011274324730038643, + 0.006366105750203133, + 0.0062841810286045074, + 0.055008672177791595, + 0.008437949232757092, + 0.017338911071419716, + 0.03654823079705238, + 0.007672959938645363, + 0.0034782811999320984, + 0.0027292247395962477, + 0.002726165810599923, + 0.015719374641776085, + 0.002829950302839279, + 0.0049240728840231895, + 0.006878510117530823, + 0.003397291759029031, + 0.00482801953330636, + 0.0031687715090811253, + 0.0038476171903312206, + 0.01957426592707634, + 0.004398476332426071, + 0.004665368236601353, + 0.021027496084570885, + 0.005125116091221571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.0056923432275652885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596086978912354 + ], + "opposite_logprob_tokens": [ + -3.8955960273742676 + ], + "logprob_reward": -0.020596086978912354, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596086978912354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12082359939813614, + 0.1791602224111557, + 0.1580961048603058, + 0.02838168479502201, + 0.3472476303577423, + 0.023111509159207344, + 0.010096740908920765, + 0.047539010643959045, + 0.02507025934755802, + 0.0604732446372509 + ], + "credit_map": [ + 0.009518961422145367, + 0.00511181028559804, + 0.007700113579630852, + 0.04398363083600998, + 0.005140624474734068, + 0.029367277398705482, + 0.007367465645074844, + 0.012633712962269783, + 0.012196331284940243, + 0.006577113643288612, + 0.006888983305543661, + 0.07247192412614822, + 0.004811454098671675, + 0.04734707623720169, + 0.007395884953439236, + 0.021471457555890083, + 0.011366447433829308, + 0.006844212766736746, + 0.007612138520926237, + 0.06017490103840828, + 0.005230611190199852, + 0.04519331827759743, + 0.006188190542161465, + 0.015486283227801323, + 0.002358947414904833, + 0.0019004687201231718, + 0.00279849860817194, + 0.0072285812348127365, + 0.0023124953731894493, + 0.0037180157378315926, + 0.003881295910105109, + 0.004183382261544466, + 0.030513819307088852, + 0.024843595921993256, + 0.03194505348801613, + 0.09202250093221664, + 0.02656739018857479, + 0.05428871512413025, + 0.039545319974422455, + 0.0475212037563324, + 0.0022587382700294256, + 0.0025634421035647392, + 0.002102126367390156, + 0.004955443553626537, + 0.0020901751704514027, + 0.00398856308311224, + 0.002510243095457554, + 0.0026427789125591516, + 0.0012705076951533556, + 0.0010155271738767624, + 0.0011661628959700465, + 0.0016432353295385838, + 0.0014853955945000052, + 0.0009871155489236116, + 0.0011448666919022799, + 0.0013839300954714417, + 0.0035805231891572475, + 0.0022263210266828537, + 0.0020074513740837574, + 0.017850840464234352, + 0.0028276622761040926, + 0.006060367450118065, + 0.010513355024158955, + 0.0024724877439439297, + 0.0019663411658257246, + 0.0016858067829161882, + 0.001762297237291932, + 0.006521810777485371, + 0.0019472616259008646, + 0.002259979024529457, + 0.007022491190582514, + 0.0019042687490582466, + 0.0049111866392195225, + 0.003507944056764245, + 0.0033686442766338587, + 0.01686899922788143, + 0.003793498734012246, + 0.004589312244206667, + 0.01716727763414383, + 0.0062663801945745945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.020596086978912354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537943199276924 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537943199276924, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537943199276924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12205880135297775, + 0.41516974568367004, + 0.04814440757036209, + 0.14092044532299042, + 0.048477500677108765, + 0.02192673832178116, + 0.012773456051945686, + 0.07841791957616806, + 0.0381208099424839, + 0.07399013638496399 + ], + "credit_map": [ + 0.008615938015282154, + 0.005834298208355904, + 0.005322395358234644, + 0.04454885050654411, + 0.004840662237256765, + 0.034859467297792435, + 0.005587421823292971, + 0.01244976557791233, + 0.03000853769481182, + 0.015844039618968964, + 0.016967155039310455, + 0.1699298918247223, + 0.012192780151963234, + 0.10771263390779495, + 0.01707071252167225, + 0.045444000512361526, + 0.003468783339485526, + 0.0022490019910037518, + 0.002373376162722707, + 0.019159043207764626, + 0.002375319367274642, + 0.011043586768209934, + 0.0016621379181742668, + 0.005813159514218569, + 0.01001135352998972, + 0.004833776969462633, + 0.004860075656324625, + 0.05817786604166031, + 0.004660678096115589, + 0.037609267979860306, + 0.0042125023901462555, + 0.016554927453398705, + 0.0049104331992566586, + 0.0033669027034193277, + 0.0038301441818475723, + 0.006863183341920376, + 0.004353516735136509, + 0.004238408990204334, + 0.013281455263495445, + 0.007633457891643047, + 0.002111338544636965, + 0.0016264759469777346, + 0.0017387213883921504, + 0.006216157227754593, + 0.0021149935200810432, + 0.003976420499384403, + 0.002133431611582637, + 0.0020092001650482416, + 0.0015497453277930617, + 0.0009957869770005345, + 0.001456665457226336, + 0.002386395586654544, + 0.0016550567233934999, + 0.0012798351235687733, + 0.0013902821810916066, + 0.0020596880931407213, + 0.005947162862867117, + 0.003943826537579298, + 0.0035213164519518614, + 0.027532052248716354, + 0.005624060984700918, + 0.009460723027586937, + 0.01792162097990513, + 0.00446716183796525, + 0.002815966960042715, + 0.00256469938904047, + 0.0021671124268323183, + 0.012047536671161652, + 0.0034254505299031734, + 0.004778311122208834, + 0.008008361794054508, + 0.0023133717477321625, + 0.0051541500724852085, + 0.003015560330823064, + 0.0035225916653871536, + 0.02378186210989952, + 0.00455116294324398, + 0.0048957993276417255, + 0.023496616631746292, + 0.005572396330535412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.012537943199276924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294490844011307 + ], + "opposite_logprob_tokens": [ + -2.2379448413848877 + ], + "logprob_reward": -0.11294490844011307, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294490844011307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10407925397157669, + 0.3449028730392456, + 0.05494764447212219, + 0.04087807610630989, + 0.1004871129989624, + 0.031593263149261475, + 0.02595566213130951, + 0.1875118613243103, + 0.0805605947971344, + 0.02908368594944477 + ], + "credit_map": [ + 0.007312968838959932, + 0.004528510384261608, + 0.0042076800018548965, + 0.041727058589458466, + 0.00431146239861846, + 0.024846412241458893, + 0.004732551518827677, + 0.01241261512041092, + 0.02389632910490036, + 0.011832295916974545, + 0.012013389728963375, + 0.14310473203659058, + 0.009967743419110775, + 0.09433566778898239, + 0.00965092983096838, + 0.04010177403688431, + 0.004254048690199852, + 0.0027758493088185787, + 0.002648628782480955, + 0.022243782877922058, + 0.002056064084172249, + 0.01263481006026268, + 0.0016543748788535595, + 0.006680089049041271, + 0.002854231046512723, + 0.0017991227796301246, + 0.0016127960989251733, + 0.01577811688184738, + 0.0016728220507502556, + 0.011226697824895382, + 0.0018395696533843875, + 0.004094719421118498, + 0.009983516298234463, + 0.0065312739461660385, + 0.007856943644583225, + 0.01581023447215557, + 0.008325206115841866, + 0.008887729607522488, + 0.025969570502638817, + 0.01712264120578766, + 0.0026364102959632874, + 0.002424933249130845, + 0.0018879380077123642, + 0.008854382671415806, + 0.0027874454390257597, + 0.005969290155917406, + 0.0036803213879466057, + 0.003352544968947768, + 0.002743433928117156, + 0.0019455578876659274, + 0.0018425239250063896, + 0.006244300398975611, + 0.0024541006423532963, + 0.002105731749907136, + 0.005895367357879877, + 0.0027246444951742887, + 0.014806731604039669, + 0.008769646286964417, + 0.00826113298535347, + 0.06693427264690399, + 0.012053137645125389, + 0.021042685955762863, + 0.044464074075222015, + 0.01118016242980957, + 0.006004132330417633, + 0.004922601394355297, + 0.004145407117903233, + 0.02658003382384777, + 0.006694272626191378, + 0.010016745887696743, + 0.016869300976395607, + 0.005328105762600899, + 0.0021295042242854834, + 0.0016703663859516382, + 0.001170663395896554, + 0.007685545366257429, + 0.0029320651665329933, + 0.002275392645969987, + 0.007691457401961088, + 0.0035286915954202414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.11294490844011307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 figure's head resting on a table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the table surface made of dark, polished wood with visible swirling grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single white paper visible in the foreground lying flat on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper gently lift and flutter with each breath before settling back down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm, low-angle light creating rich browns and soft golds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011801904765889049 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011801904765889049, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011801904765889049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31148287653923035, + 0.34717798233032227, + 0.03891724348068237, + 0.09702553600072861, + 0.05308844521641731, + 0.041990771889686584, + 0.02059601992368698, + 0.03792136535048485, + 0.008211127482354641, + 0.04358861222863197 + ], + "credit_map": [ + 0.021448804065585136, + 0.012008260935544968, + 0.01214427687227726, + 0.12873657047748566, + 0.008596992120146751, + 0.08545726537704468, + 0.00967632606625557, + 0.03341442719101906, + 0.024690601974725723, + 0.011381514370441437, + 0.013251712545752525, + 0.14583614468574524, + 0.008785588666796684, + 0.09085100144147873, + 0.010271825827658176, + 0.04210963100194931, + 0.003063597483560443, + 0.00395370926707983, + 0.0038887953851372004, + 0.010119335725903511, + 0.002634407952427864, + 0.007521704770624638, + 0.0034014605917036533, + 0.0043342383578419685, + 0.007401397917419672, + 0.004063532687723637, + 0.003940627444535494, + 0.038870669901371, + 0.0034012494143098593, + 0.02395813725888729, + 0.004252004902809858, + 0.011137925088405609, + 0.005204050336033106, + 0.0043221511878073215, + 0.005707009229809046, + 0.012558909133076668, + 0.004213117063045502, + 0.007368776481598616, + 0.007049189414829016, + 0.006665243301540613, + 0.003502362407743931, + 0.003267377382144332, + 0.0045056636445224285, + 0.006685416214168072, + 0.003232297021895647, + 0.003997587598860264, + 0.007907897233963013, + 0.008892173878848553, + 0.001878754235804081, + 0.0010727146873250604, + 0.001360634109005332, + 0.0070365010760724545, + 0.001349394558928907, + 0.0021103473845869303, + 0.004315248224884272, + 0.0014724272768944502, + 0.002635762793943286, + 0.002081406069919467, + 0.001771090435795486, + 0.013265705667436123, + 0.002870226977393031, + 0.00506409490481019, + 0.008398463018238544, + 0.0018346176948398352, + 0.0010686102323234081, + 0.0007022538920864463, + 0.0009598170290701091, + 0.0020069130696356297, + 0.0006211060681380332, + 0.0006662854575552046, + 0.0012451180955395103, + 0.0009410246857441962, + 0.003372561652213335, + 0.0022876840084791183, + 0.0022900474723428488, + 0.013372243382036686, + 0.0030446494929492474, + 0.0028998262714594603, + 0.012691012583673, + 0.0036305920220911503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0011801904765889049 + } + ], + "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/0000131/group_7/index_tensor([7621, 7621])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_tensor([7621, 7621])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..90df0138b1008407b4e740d03005cdae61951a5b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_tensor([7621, 7621])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot, the camera moving smoothly sideways, hugging the water's surface to keep pace with the subject. In the soft golden light of early morning, the vast, tranquil lake resembles a vast, unpolished silver mirror, reflecting the delicate colors of the sky and the distant, mist-shrouded mountain silhouettes. A swallow, its back feathers a deep blue-black sheen in the light, its belly a pristine, milky white, enters the frame with breathtaking speed. Its body is dramatically lowered, almost parallel to the water's surface, its streamlined form a picture of perfect aerodynamics. With powerful and graceful flapping of its long, pointed wings, each downward stroke, the tips of the wings slice into the water with exquisite precision and lightness, tandem, instantly carving brief, delicate V-shaped ripples across the glassy surface. These subtle ripples quickly spread outward, then swallowed by the lake's tranquility, leaving only a barely visible trail behind the swallow. The entire picture is full of dynamic beauty in tranquility, with soft light and shadow, and the tones are mainly cool lake blue and warm morning light, creating a poetic atmosphere that is both full of vitality and extremely peaceful and breathtaking.", + "seed": 16736, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16736_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": -0.10138118020258843, + "vlm_logprob": -0.10138118020258843, + "vlm_yes_no_margin": 3.449999952316284, + "vlm_token_credit": -0.10138118020258843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3476400673389435 + ], + "opposite_logprob_tokens": [ + -1.472640037536621 + ], + "logprob_reward": -0.3476400673389435, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3476400673389435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3134450912475586, + 0.035598620772361755, + 0.049370963126420975, + 0.09412030875682831, + 0.09289278090000153, + 0.0722808688879013, + 0.1180882602930069, + 0.10923551768064499, + 0.07047630101442337, + 0.0444912388920784 + ], + "credit_map": [ + 0.031218362972140312, + 0.030778642743825912, + 0.057472940534353256, + 0.043024491518735886, + 0.01883663609623909, + 0.024899261072278023, + 0.09009027481079102, + 0.017124509438872337, + 0.003483472391963005, + 0.0040110014379024506, + 0.005039590410888195, + 0.00466172257438302, + 0.004101231694221497, + 0.004519258625805378, + 0.005174154415726662, + 0.004608190152794123, + 0.006670190021395683, + 0.006297963671386242, + 0.0055121080949902534, + 0.005896396469324827, + 0.007798774633556604, + 0.005210929550230503, + 0.006230560597032309, + 0.00575404055416584, + 0.008790282532572746, + 0.007991315796971321, + 0.00971805490553379, + 0.013574901036918163, + 0.013478325679898262, + 0.007236097473651171, + 0.022165002301335335, + 0.011166329495608807, + 0.009790592826902866, + 0.012414923869073391, + 0.012371651828289032, + 0.010882030241191387, + 0.015035383403301239, + 0.01151095237582922, + 0.008781741373240948, + 0.012105506844818592, + 0.006441952660679817, + 0.008103050291538239, + 0.006906080991029739, + 0.007190926466137171, + 0.01713603362441063, + 0.00991983525454998, + 0.008559691719710827, + 0.008023295551538467, + 0.01381066907197237, + 0.026384059339761734, + 0.01291581615805626, + 0.008489207364618778, + 0.013017443008720875, + 0.012348298914730549, + 0.014102412387728691, + 0.017020361497998238, + 0.012332088313996792, + 0.017988549545407295, + 0.014176096767187119, + 0.009120957925915718, + 0.016763757914304733, + 0.015416296198964119, + 0.010902312584221363, + 0.012535462155938148, + 0.0065519437193870544, + 0.004588201176375151, + 0.005183979868888855, + 0.004513554275035858, + 0.00503019243478775, + 0.008412088267505169, + 0.01776628941297531, + 0.01843004673719406, + 0.004799745976924896, + 0.00980792660266161, + 0.0024197108577936888, + 0.00269767246209085, + 0.005885678343474865, + 0.01153353787958622, + 0.003682236187160015, + 0.0036647312808781862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.3476400673389435 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298345774412155 + ], + "opposite_logprob_tokens": [ + -2.237983465194702 + ], + "logprob_reward": -0.11298345774412155, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298345774412155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18974944949150085, + 0.07037343084812164, + 0.05709575489163399, + 0.1277998387813568, + 0.11434359103441238, + 0.12748222053050995, + 0.07772588729858398, + 0.0854877308011055, + 0.0782696083188057, + 0.0716724619269371 + ], + "credit_map": [ + 0.018683485686779022, + 0.01820305362343788, + 0.053960300981998444, + 0.014637811109423637, + 0.011058240197598934, + 0.010913434438407421, + 0.05076015368103981, + 0.011532974429428577, + 0.005299756769090891, + 0.006635935045778751, + 0.009756037034094334, + 0.016089485958218575, + 0.005591651424765587, + 0.012843821197748184, + 0.006724142469465733, + 0.0074326107278466225, + 0.003983371891081333, + 0.009221220389008522, + 0.008977086283266544, + 0.010874759405851364, + 0.005422390531748533, + 0.005136185325682163, + 0.008180974051356316, + 0.005299773067235947, + 0.00877747405320406, + 0.009425572119653225, + 0.013656947761774063, + 0.029845261946320534, + 0.012752886861562729, + 0.01253175362944603, + 0.024528980255126953, + 0.01628097891807556, + 0.009806123562157154, + 0.016088806092739105, + 0.014679960906505585, + 0.012714026495814323, + 0.017562439665198326, + 0.01243302971124649, + 0.011629018001258373, + 0.019430197775363922, + 0.010790755972266197, + 0.019744666293263435, + 0.014850410632789135, + 0.015140672214329243, + 0.025302046909928322, + 0.021833093836903572, + 0.010468650609254837, + 0.009351922199130058, + 0.01199288945645094, + 0.014104915782809258, + 0.009511267766356468, + 0.00534391263499856, + 0.010984324850142002, + 0.009295010939240456, + 0.007763230707496405, + 0.008730337955057621, + 0.005943900439888239, + 0.01583687588572502, + 0.010248735547065735, + 0.007236141245812178, + 0.013465936295688152, + 0.010324626229703426, + 0.009892312809824944, + 0.01253920141607523, + 0.004910716321319342, + 0.007181401364505291, + 0.008082743734121323, + 0.005767250899225473, + 0.004721591714769602, + 0.004701963625848293, + 0.023425882682204247, + 0.01947806589305401, + 0.007612088229507208, + 0.016984442248940468, + 0.004683229140937328, + 0.004282769281417131, + 0.010625665076076984, + 0.011348004452884197, + 0.0076522123999893665, + 0.00848405808210373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.11298345774412155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821867562830448 + ], + "opposite_logprob_tokens": [ + -4.018218517303467 + ], + "logprob_reward": -0.01821867562830448, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821867562830448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11020997166633606, + 0.022227520123124123, + 0.03618466481566429, + 0.053482212126255035, + 0.03862916678190231, + 0.18984006345272064, + 0.12896598875522614, + 0.28713664412498474, + 0.0756412073969841, + 0.05768253281712532 + ], + "credit_map": [ + 0.009670714847743511, + 0.008727599866688251, + 0.027112698182463646, + 0.016693515703082085, + 0.010045916773378849, + 0.018403643742203712, + 0.01192064955830574, + 0.007635240908712149, + 0.003012679750099778, + 0.00178715109359473, + 0.0029155544470995665, + 0.0039696828462183475, + 0.0019491531420499086, + 0.002871851669624448, + 0.002316083060577512, + 0.0034053646959364414, + 0.0030478585977107286, + 0.003957739565521479, + 0.0049078455194830894, + 0.00512631656602025, + 0.006412964314222336, + 0.003979326691478491, + 0.004337902180850506, + 0.00441471254453063, + 0.0030365067068487406, + 0.0038111722096800804, + 0.010242760181427002, + 0.006877922918647528, + 0.006669262424111366, + 0.0033693930599838495, + 0.015043304301798344, + 0.004431889392435551, + 0.0041985297575592995, + 0.004801448434591293, + 0.005496656522154808, + 0.00627658748999238, + 0.004978561773896217, + 0.00405088672414422, + 0.0035159490071237087, + 0.0053105466067790985, + 0.017926380038261414, + 0.023031778633594513, + 0.021901819854974747, + 0.02258121222257614, + 0.04817328229546547, + 0.027035558596253395, + 0.012649033218622208, + 0.016540993005037308, + 0.012928115203976631, + 0.025609096512198448, + 0.013775156810879707, + 0.013413406908512115, + 0.0169513002038002, + 0.01445780973881483, + 0.016033310443162918, + 0.015797797590494156, + 0.03346848487854004, + 0.04909924045205116, + 0.0451866090297699, + 0.024115193635225296, + 0.03861279785633087, + 0.033254094421863556, + 0.02881120890378952, + 0.0345890149474144, + 0.007975956425070763, + 0.0039422838017344475, + 0.0060999151319265366, + 0.005419743712991476, + 0.004638180136680603, + 0.004323319066315889, + 0.02111934870481491, + 0.02212246134877205, + 0.007501064334064722, + 0.010645152069628239, + 0.0034091032575815916, + 0.004459355026483536, + 0.0054372744634747505, + 0.012556343339383602, + 0.007245532236993313, + 0.006428703665733337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.01821867562830448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716808183118701 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716808183118701, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716808183118701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18146182596683502, + 0.05041281506419182, + 0.05237344279885292, + 0.13432566821575165, + 0.052328817546367645, + 0.18712353706359863, + 0.08158053457736969, + 0.11081601679325104, + 0.07954788208007812, + 0.07002944499254227 + ], + "credit_map": [ + 0.013294135220348835, + 0.011127904988825321, + 0.02817200869321823, + 0.03711879253387451, + 0.013270541094243526, + 0.025158455595374107, + 0.03230558708310127, + 0.021014386788010597, + 0.0065857539884746075, + 0.00708338338881731, + 0.006696485448628664, + 0.009741139598190784, + 0.003995687235146761, + 0.0056513831950724125, + 0.005011796485632658, + 0.005647188052535057, + 0.004647924564778805, + 0.01063940953463316, + 0.00815509632229805, + 0.008149843662977219, + 0.005692544858902693, + 0.004922842141240835, + 0.005493828095495701, + 0.004671956412494183, + 0.008650300092995167, + 0.010206349194049835, + 0.015412221662700176, + 0.030295638367533684, + 0.015130466781556606, + 0.011343911290168762, + 0.024553366005420685, + 0.01873341202735901, + 0.005950475577265024, + 0.005418552551418543, + 0.0062268683686852455, + 0.00552318524569273, + 0.010097428224980831, + 0.00732045853510499, + 0.004564984235912561, + 0.00722686480730772, + 0.020314743742346764, + 0.026577165350317955, + 0.018342304974794388, + 0.017385168001055717, + 0.030550610274076462, + 0.038816530257463455, + 0.015787580981850624, + 0.019349422305822372, + 0.010353508405387402, + 0.017127012833952904, + 0.008435674011707306, + 0.006715723779052496, + 0.011723932810127735, + 0.007881821133196354, + 0.009103335440158844, + 0.010239535011351109, + 0.0091856699436903, + 0.016115497797727585, + 0.017119526863098145, + 0.011936886236071587, + 0.014629017561674118, + 0.01145431213080883, + 0.015368652530014515, + 0.015006461180746555, + 0.012530509382486343, + 0.009075585752725601, + 0.010776159353554249, + 0.009196233935654163, + 0.008434746414422989, + 0.009433364495635033, + 0.007013293448835611, + 0.013087996281683445, + 0.011747532524168491, + 0.016271179541945457, + 0.0034162879455834627, + 0.004504360258579254, + 0.006366024725139141, + 0.01423556823283434, + 0.006567179691046476, + 0.006921309977769852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.001716808183118701 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634689211845398 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.02634689211845398, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634689211845398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1644025295972824, + 0.060118209570646286, + 0.08023080974817276, + 0.03898043930530548, + 0.04999079182744026, + 0.1018126904964447, + 0.0726630836725235, + 0.2658042013645172, + 0.1088276356458664, + 0.057169608771800995 + ], + "credit_map": [ + 0.010232825763523579, + 0.009004260413348675, + 0.05198683962225914, + 0.02862006425857544, + 0.012591634877026081, + 0.020887045189738274, + 0.022004293277859688, + 0.009075573645532131, + 0.005407822318375111, + 0.007319977506995201, + 0.010021578520536423, + 0.015466178767383099, + 0.004529732745140791, + 0.007185713388025761, + 0.004987995605915785, + 0.005199210252612829, + 0.007675270549952984, + 0.018115712329745293, + 0.010904202237725258, + 0.011310560628771782, + 0.010904346592724323, + 0.0059759728610515594, + 0.008207984268665314, + 0.0071367681957781315, + 0.003644447773694992, + 0.0032600981649011374, + 0.00492077274248004, + 0.006376440171152353, + 0.004070765804499388, + 0.003234430681914091, + 0.007828352972865105, + 0.00564513448625803, + 0.004652251023799181, + 0.0060589988715946674, + 0.005937039852142334, + 0.0050727603957057, + 0.007867579348385334, + 0.005988507065922022, + 0.005092031788080931, + 0.00932161882519722, + 0.00997810997068882, + 0.010221585631370544, + 0.011316046118736267, + 0.009872066788375378, + 0.02687920071184635, + 0.016526320949196815, + 0.00804169848561287, + 0.008977657184004784, + 0.00750234629958868, + 0.014474734663963318, + 0.0068908375687897205, + 0.0067273639142513275, + 0.009190686978399754, + 0.007951430045068264, + 0.00954689085483551, + 0.010378791950643063, + 0.028013957664370537, + 0.046674881130456924, + 0.042698170989751816, + 0.018779635429382324, + 0.03740566596388817, + 0.036321356892585754, + 0.025429226458072662, + 0.030481290072202682, + 0.012758004479110241, + 0.004587939474731684, + 0.008914519101381302, + 0.003845629747956991, + 0.007015279959887266, + 0.0059758867137134075, + 0.03445611521601677, + 0.03127426654100418, + 0.009234603494405746, + 0.01567712426185608, + 0.003120198380202055, + 0.003498187055811286, + 0.005521484650671482, + 0.008847629651427269, + 0.005485116504132748, + 0.005785265006124973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02634689211845398 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, the camera moving smoothly sideways, hugging the water's surface to keep pace with the subject. In the soft golden light of early morning, the vast, tranquil lake resembles a vast, unpolished silver mirror, reflecting the delicate colors of the sky and the distant, mist-shrouded mountain silhouettes. A swallow, its back feathers a deep blue-black sheen in the light, its belly a pristine, milky white, enters the frame with breathtaking speed. Its body is dramatically lowered, almost parallel to the water's surface, its streamlined form a picture of perfect aerodynamics. With powerful and graceful flapping of its long, pointed wings, each downward stroke, the tips of the wings slice into the water with exquisite precision and lightness, tandem, instantly carving brief, delicate V-shaped ripples across the glassy surface. These subtle ripples quickly spread outward, then swallowed by the lake's tranquility, leaving only a barely visible trail behind the swallow. The entire picture is full of dynamic beauty in tranquility, with soft light and shadow, and the tones are mainly cool lake blue and warm morning light, creating a poetic atmosphere that is both full of vitality and extremely peaceful and breathtaking.", + "seed": 16737, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16737_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": -0.09719796846620739, + "vlm_logprob": -0.09719796846620739, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.09719796846620739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26817089319229126 + ], + "opposite_logprob_tokens": [ + -1.6431708335876465 + ], + "logprob_reward": -0.26817089319229126, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.26817089319229126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3680480420589447, + 0.21362976729869843, + 0.04559306055307388, + 0.050404202193021774, + 0.10766314715147018, + 0.04161008447408676, + 0.03730626031756401, + 0.07381642609834671, + 0.04257165640592575, + 0.019357457756996155 + ], + "credit_map": [ + 0.0296450424939394, + 0.09953434020280838, + 0.022375859320163727, + 0.02454334869980812, + 0.09026426076889038, + 0.05225454643368721, + 0.033562999218702316, + 0.015867626294493675, + 0.007146759424358606, + 0.06426721811294556, + 0.008997861295938492, + 0.010814276523888111, + 0.026568813249468803, + 0.011894297786056995, + 0.06442853808403015, + 0.01951199397444725, + 0.004388967528939247, + 0.005784106906503439, + 0.007161557674407959, + 0.008202512748539448, + 0.004645496606826782, + 0.004556561354547739, + 0.00564946373924613, + 0.0052043963223695755, + 0.004611215554177761, + 0.01175119262188673, + 0.005022563971579075, + 0.006868300028145313, + 0.006240558344870806, + 0.00630780216306448, + 0.003649261314421892, + 0.00595331285148859, + 0.012742140330374241, + 0.019552476704120636, + 0.011139470152556896, + 0.013306799344718456, + 0.017464328557252884, + 0.01087751891463995, + 0.011545268818736076, + 0.011035150848329067, + 0.004918832331895828, + 0.007393333595246077, + 0.003825669176876545, + 0.0036651857662945986, + 0.00452776812016964, + 0.006342364940792322, + 0.0046021281741559505, + 0.006334799341857433, + 0.0038938219659030437, + 0.005560855846852064, + 0.0033202585764229298, + 0.0038828207179903984, + 0.0064004333689808846, + 0.005825221072882414, + 0.003592893248423934, + 0.004829955752938986, + 0.009953947737812996, + 0.02206256054341793, + 0.005095773376524448, + 0.008217023685574532, + 0.0059805321507155895, + 0.013996671885251999, + 0.00381659553386271, + 0.004693320021033287, + 0.00485026091337204, + 0.005892137531191111, + 0.0028080095071345568, + 0.0042236181907355785, + 0.004175265319645405, + 0.012041637673974037, + 0.0034941521007567644, + 0.005086575634777546, + 0.0022209433373063803, + 0.0028933007270097733, + 0.0014123719884082675, + 0.0023164006415754557, + 0.003922054078429937, + 0.0030319439247250557, + 0.0019340235739946365, + 0.001626421813853085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.26817089319229126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042152047157288 + ], + "opposite_logprob_tokens": [ + -1.9104214906692505 + ], + "logprob_reward": -0.16042152047157288, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042152047157288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08858852833509445, + 0.19892127811908722, + 0.04650687426328659, + 0.038807276636362076, + 0.18101662397384644, + 0.17494769394397736, + 0.07968290150165558, + 0.09745731949806213, + 0.0519726537168026, + 0.04209887608885765 + ], + "credit_map": [ + 0.013443078845739365, + 0.015542460605502129, + 0.01055279839783907, + 0.012280134484171867, + 0.009087991900742054, + 0.0091858496889472, + 0.007768444716930389, + 0.010727767832577229, + 0.006206299643963575, + 0.02620491199195385, + 0.00935731828212738, + 0.011185357347130775, + 0.025346340611577034, + 0.03591243922710419, + 0.06394392251968384, + 0.020764682441949844, + 0.0037618547212332487, + 0.006014470476657152, + 0.006699446588754654, + 0.009401779621839523, + 0.00422775698825717, + 0.005740748252719641, + 0.005744927562773228, + 0.0049158912152051926, + 0.004364511929452419, + 0.007027089595794678, + 0.004546881653368473, + 0.0056219943799078465, + 0.005318982060998678, + 0.00410141097381711, + 0.004381921142339706, + 0.0034444862976670265, + 0.01890791766345501, + 0.03440297022461891, + 0.020608846098184586, + 0.02525612711906433, + 0.022157883271574974, + 0.018587913364171982, + 0.020740848034620285, + 0.02035410702228546, + 0.021998552605509758, + 0.01973761059343815, + 0.02572442963719368, + 0.013204765506088734, + 0.013385478407144547, + 0.02297389507293701, + 0.016226928681135178, + 0.04169603809714317, + 0.009380056522786617, + 0.011163824237883091, + 0.00698227621614933, + 0.009107650257647038, + 0.011241476982831955, + 0.016225147992372513, + 0.007588766049593687, + 0.007993708364665508, + 0.011911194771528244, + 0.021346183493733406, + 0.007835911586880684, + 0.01216797810047865, + 0.014300046488642693, + 0.014211215078830719, + 0.00871946755796671, + 0.006965317763388157, + 0.006505353841930628, + 0.007177895400673151, + 0.004506631754338741, + 0.005722059868276119, + 0.005586163140833378, + 0.013523394241929054, + 0.003815253498032689, + 0.0051359026692807674, + 0.00416923314332962, + 0.005746678914874792, + 0.002888691145926714, + 0.004103699699044228, + 0.009560608305037022, + 0.007998782210052013, + 0.004034563899040222, + 0.0035966180730611086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.16042152047157288 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609875075519085 + ], + "opposite_logprob_tokens": [ + -4.141098976135254 + ], + "logprob_reward": -0.01609875075519085, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609875075519085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09178558737039566, + 0.18769726157188416, + 0.10260435938835144, + 0.04946303740143776, + 0.3110470473766327, + 0.02998231165111065, + 0.025509174913167953, + 0.13649052381515503, + 0.017688246443867683, + 0.047732412815093994 + ], + "credit_map": [ + 0.014069711789488792, + 0.01371525228023529, + 0.012074939906597137, + 0.01245109923183918, + 0.011519943363964558, + 0.00851671863347292, + 0.009261813946068287, + 0.010176101699471474, + 0.004956857301294804, + 0.05759035423398018, + 0.007603379897773266, + 0.006488027051091194, + 0.010793168097734451, + 0.07134487479925156, + 0.01961360312998295, + 0.009306983090937138, + 0.009019916877150536, + 0.012737998738884926, + 0.017634300515055656, + 0.02093348652124405, + 0.0093772541731596, + 0.009910114109516144, + 0.011014397256076336, + 0.011976897716522217, + 0.00494431285187602, + 0.009506713598966599, + 0.004626585636287928, + 0.008354101330041885, + 0.0065923710353672504, + 0.006711004301905632, + 0.003819281468167901, + 0.004908663686364889, + 0.03348866105079651, + 0.043189886957407, + 0.037484318017959595, + 0.0444120392203331, + 0.042926687747240067, + 0.030920404940843582, + 0.03722050040960312, + 0.04140453040599823, + 0.004678167402744293, + 0.0037202974781394005, + 0.002674939576536417, + 0.0031120493076741695, + 0.004105363041162491, + 0.005042669828981161, + 0.0035852158907800913, + 0.0030636063311249018, + 0.002817289438098669, + 0.004325924441218376, + 0.0022925762459635735, + 0.0023646149784326553, + 0.003435820573940873, + 0.005384691059589386, + 0.0025822652969509363, + 0.0023059933446347713, + 0.015640733763575554, + 0.031226905062794685, + 0.009196990169584751, + 0.011523708701133728, + 0.021050943061709404, + 0.026984509080648422, + 0.012284225784242153, + 0.008582506328821182, + 0.0013733054511249065, + 0.002022160217165947, + 0.002135147573426366, + 0.004306238144636154, + 0.0015354423085227609, + 0.002228816971182823, + 0.0020574042573571205, + 0.0020297325681895018, + 0.0037401937879621983, + 0.005736540071666241, + 0.0038690390065312386, + 0.003681489499285817, + 0.00833081267774105, + 0.006373788695782423, + 0.003869312582537532, + 0.012131234630942345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.01609875075519085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031985098030418158 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031985098030418158, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031985098030418158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12826479971408844, + 0.10617333650588989, + 0.05221998691558838, + 0.06763529777526855, + 0.25208672881126404, + 0.0462832897901535, + 0.08780449628829956, + 0.18408483266830444, + 0.03365231677889824, + 0.041794952005147934 + ], + "credit_map": [ + 0.012761469930410385, + 0.033545974642038345, + 0.009638241492211819, + 0.010251889936625957, + 0.026957543566823006, + 0.015571901574730873, + 0.011685747653245926, + 0.007852012291550636, + 0.004756648559123278, + 0.03812432661652565, + 0.006182560697197914, + 0.004648255184292793, + 0.011679508723318577, + 0.018109122291207314, + 0.01661648415029049, + 0.0060564251616597176, + 0.0055151828564703465, + 0.007578540127724409, + 0.005962485913187265, + 0.009210126474499702, + 0.005513340700417757, + 0.007116121705621481, + 0.006643395405262709, + 0.004680787678807974, + 0.006532975006848574, + 0.015300248749554157, + 0.005865559447556734, + 0.010243840515613556, + 0.009030145592987537, + 0.007203899323940277, + 0.005760380532592535, + 0.0076982406899333, + 0.028429634869098663, + 0.03393137827515602, + 0.031364038586616516, + 0.03450446203351021, + 0.0325046144425869, + 0.024043696001172066, + 0.03241507709026337, + 0.034893810749053955, + 0.005263053812086582, + 0.00624921964481473, + 0.0060631996020674706, + 0.004300243686884642, + 0.0038867697585374117, + 0.0061389850452542305, + 0.005793130025267601, + 0.008588684722781181, + 0.009933510795235634, + 0.01631665788590908, + 0.006929384078830481, + 0.008153945207595825, + 0.012787326239049435, + 0.016149889677762985, + 0.008700955659151077, + 0.008832825347781181, + 0.02282947115600109, + 0.0449042022228241, + 0.014953330159187317, + 0.015193129889667034, + 0.02330472134053707, + 0.03610091656446457, + 0.015062354505062103, + 0.01173669844865799, + 0.0038698448333889246, + 0.004420834593474865, + 0.002466668142005801, + 0.003764830529689789, + 0.0038202288560569286, + 0.007304502185434103, + 0.004189611412584782, + 0.0038157934322953224, + 0.0039925104938447475, + 0.006268952041864395, + 0.002842081245034933, + 0.004206457175314426, + 0.009778717532753944, + 0.007829618640244007, + 0.003813484450802207, + 0.0030631241388618946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.0031985098030418158 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038100168108940125 + ], + "opposite_logprob_tokens": [ + -3.288100242614746 + ], + "logprob_reward": -0.038100168108940125, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038100168108940125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15574826300144196, + 0.3611094355583191, + 0.05081930384039879, + 0.024976549670100212, + 0.15623541176319122, + 0.04018603637814522, + 0.05750830098986626, + 0.09397057443857193, + 0.03446788340806961, + 0.024978237226605415 + ], + "credit_map": [ + 0.01835775002837181, + 0.03549681603908539, + 0.01153570506721735, + 0.011443419381976128, + 0.036646537482738495, + 0.02006593532860279, + 0.013879789039492607, + 0.008322316221892834, + 0.010206173174083233, + 0.1422184705734253, + 0.007013488560914993, + 0.01121291983872652, + 0.029883356764912605, + 0.07553479075431824, + 0.06529393047094345, + 0.01974629983305931, + 0.005318391136825085, + 0.005804191809147596, + 0.006969713140279055, + 0.008719094097614288, + 0.005112337414175272, + 0.007796150166541338, + 0.0058132354170084, + 0.005286190193146467, + 0.0026348375249654055, + 0.005253559444099665, + 0.0027147457003593445, + 0.0036972924135625362, + 0.0031113596633076668, + 0.002739853924140334, + 0.002332644537091255, + 0.002492256462574005, + 0.025152752175927162, + 0.024869928136467934, + 0.015377652831375599, + 0.01815791241824627, + 0.020476987585425377, + 0.020503785461187363, + 0.015725797042250633, + 0.015970591455698013, + 0.005106006748974323, + 0.003951010759919882, + 0.005586646031588316, + 0.0025090142153203487, + 0.003577316179871559, + 0.005615407135337591, + 0.003589220345020294, + 0.010251414962112904, + 0.006977927405387163, + 0.010823267512023449, + 0.004621380940079689, + 0.005833714734762907, + 0.007606962695717812, + 0.009978564456105232, + 0.005857488140463829, + 0.005808994639664888, + 0.01130494475364685, + 0.018801234662532806, + 0.00609125941991806, + 0.00743819260969758, + 0.016121244058012962, + 0.018582144752144814, + 0.009705347940325737, + 0.0059262108989059925, + 0.003493842203170061, + 0.0044507537968456745, + 0.002676376374438405, + 0.004206330049782991, + 0.004253427032381296, + 0.006103288382291794, + 0.005414493847638369, + 0.0038693719543516636, + 0.0020228291396051645, + 0.0032210496719926596, + 0.0020006804261356592, + 0.0032444903627038, + 0.004937502555549145, + 0.0036705362144857645, + 0.0025916611775755882, + 0.003289487911388278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.038100168108940125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, the camera moving smoothly sideways, hugging the water's surface to keep pace with the subject. In the soft golden light of early morning, the vast, tranquil lake resembles a vast, unpolished silver mirror, reflecting the delicate colors of the sky and the distant, mist-shrouded mountain silhouettes. A swallow, its back feathers a deep blue-black sheen in the light, its belly a pristine, milky white, enters the frame with breathtaking speed. Its body is dramatically lowered, almost parallel to the water's surface, its streamlined form a picture of perfect aerodynamics. With powerful and graceful flapping of its long, pointed wings, each downward stroke, the tips of the wings slice into the water with exquisite precision and lightness, tandem, instantly carving brief, delicate V-shaped ripples across the glassy surface. These subtle ripples quickly spread outward, then swallowed by the lake's tranquility, leaving only a barely visible trail behind the swallow. The entire picture is full of dynamic beauty in tranquility, with soft light and shadow, and the tones are mainly cool lake blue and warm morning light, creating a poetic atmosphere that is both full of vitality and extremely peaceful and breathtaking.", + "seed": 16738, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16738_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": -0.15113140798639507, + "vlm_logprob": -0.15113140798639507, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.15113140798639507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5229936838150024 + ], + "opposite_logprob_tokens": [ + -1.0229936838150024 + ], + "logprob_reward": -0.5229936838150024, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5229936838150024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2943565249443054, + 0.05847657099366188, + 0.05345494672656059, + 0.03003068082034588, + 0.08885208517313004, + 0.09127062559127808, + 0.09367222338914871, + 0.06430664658546448, + 0.05449840798974037, + 0.1710812747478485 + ], + "credit_map": [ + 0.007100718095898628, + 0.012198776006698608, + 0.011863376945257187, + 0.12213806062936783, + 0.008523451164364815, + 0.005986764095723629, + 0.11677655577659607, + 0.009768832474946976, + 0.0050547122955322266, + 0.007462720386683941, + 0.00843046698719263, + 0.008306817151606083, + 0.005479198880493641, + 0.005937519017606974, + 0.006741455756127834, + 0.011063679121434689, + 0.005494122859090567, + 0.007766778115183115, + 0.008419175632297993, + 0.0069166612811386585, + 0.007987190037965775, + 0.005218653008341789, + 0.006701280362904072, + 0.0049510798417031765, + 0.005468770861625671, + 0.0030628072563558817, + 0.0051159486174583435, + 0.004010145086795092, + 0.002745461417362094, + 0.003046781523153186, + 0.003501666011288762, + 0.0030790974851697683, + 0.006372721865773201, + 0.010830221697688103, + 0.015130756422877312, + 0.023238174617290497, + 0.007534592878073454, + 0.005685083568096161, + 0.013853040523827076, + 0.006207489408552647, + 0.011445893906056881, + 0.013981963507831097, + 0.009173394180834293, + 0.010432340204715729, + 0.01019576471298933, + 0.0119367316365242, + 0.01081241574138403, + 0.013292117975652218, + 0.010023611597716808, + 0.014108547009527683, + 0.012175893411040306, + 0.009318266063928604, + 0.012979164719581604, + 0.01646718569099903, + 0.01033202838152647, + 0.008267526514828205, + 0.01003104168921709, + 0.007218773476779461, + 0.005742671899497509, + 0.007374361157417297, + 0.008954215794801712, + 0.009156938642263412, + 0.006867365911602974, + 0.008961278945207596, + 0.00871842447668314, + 0.007219079416245222, + 0.007369691971689463, + 0.005919554736465216, + 0.006723123136907816, + 0.008441687561571598, + 0.004646840039640665, + 0.005460004322230816, + 0.026375725865364075, + 0.02406221628189087, + 0.022244803607463837, + 0.01964438334107399, + 0.016089361160993576, + 0.02135845273733139, + 0.017351768910884857, + 0.02395455352962017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.5229936838150024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163054764270782 + ], + "opposite_logprob_tokens": [ + -1.7016305923461914 + ], + "logprob_reward": -0.20163054764270782, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163054764270782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2704653739929199, + 0.04483645409345627, + 0.1041480302810669, + 0.03629831224679947, + 0.05547875538468361, + 0.03790973871946335, + 0.1574348509311676, + 0.06482304632663727, + 0.07820495963096619, + 0.1504005491733551 + ], + "credit_map": [ + 0.006459263619035482, + 0.011938856914639473, + 0.0104015301913023, + 0.1092371940612793, + 0.010945765301585197, + 0.006715376395732164, + 0.10703012347221375, + 0.007737248204648495, + 0.004618064034730196, + 0.00644296407699585, + 0.005946820136159658, + 0.005745770409703255, + 0.005747143179178238, + 0.004125891253352165, + 0.0054628499783575535, + 0.006746950559318066, + 0.010788374580442905, + 0.01010529138147831, + 0.01801062747836113, + 0.01503660436719656, + 0.010945780202746391, + 0.010765448212623596, + 0.01763886585831642, + 0.010857037268579006, + 0.004971843212842941, + 0.005925954785197973, + 0.004180664196610451, + 0.004332422744482756, + 0.003980803769081831, + 0.004021227825433016, + 0.004255341365933418, + 0.004630055744200945, + 0.005045035388320684, + 0.006911530625075102, + 0.008869516663253307, + 0.012184150516986847, + 0.0037628356367349625, + 0.0049626752734184265, + 0.007802440319210291, + 0.005940570496022701, + 0.003091296646744013, + 0.00855268444865942, + 0.003481294959783554, + 0.004608243238180876, + 0.003874144284054637, + 0.003974356222897768, + 0.0049257478676736355, + 0.005401970818638802, + 0.018065838143229485, + 0.024399416521191597, + 0.02017318829894066, + 0.013046273961663246, + 0.026880690827965736, + 0.023846570402383804, + 0.014953604899346828, + 0.016069255769252777, + 0.006553718354552984, + 0.011995420791208744, + 0.005538115743547678, + 0.006040605250746012, + 0.0144391730427742, + 0.006894405465573072, + 0.006337118800729513, + 0.007024488877505064, + 0.010384534485638142, + 0.010839736089110374, + 0.009955854155123234, + 0.010957451537251472, + 0.007937399670481682, + 0.011976896785199642, + 0.009263661690056324, + 0.0068894331343472, + 0.021950164809823036, + 0.030616354197263718, + 0.019093956798315048, + 0.017995106056332588, + 0.023487962782382965, + 0.013753991574048996, + 0.01508516725152731, + 0.008417846634984016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.20163054764270782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036353471223264933 + ], + "opposite_logprob_tokens": [ + -5.628635406494141 + ], + "logprob_reward": -0.0036353471223264933, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036353471223264933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7732390761375427, + 0.00637876195833087, + 0.02026926353573799, + 0.015642575919628143, + 0.10938780009746552, + 0.01586924120783806, + 0.0162492822855711, + 0.009314227849245071, + 0.012493270449340343, + 0.021156465634703636 + ], + "credit_map": [ + 0.007976295426487923, + 0.020546961575746536, + 0.047481417655944824, + 0.35236644744873047, + 0.015570050105452538, + 0.007631746586412191, + 0.3054791986942291, + 0.016186930239200592, + 0.0004889876581728458, + 0.0014866801211610436, + 0.0006850215140730143, + 0.0007360454183071852, + 0.001190840033814311, + 0.00046137318713590503, + 0.0006242706440389156, + 0.0007055432652123272, + 0.0021505255717784166, + 0.002263145288452506, + 0.0035712302196770906, + 0.002557766856625676, + 0.0022506839595735073, + 0.0021388649474829435, + 0.003212096868082881, + 0.0021249509882181883, + 0.003091085935011506, + 0.0023144471924751997, + 0.0025373955722898245, + 0.001957338536158204, + 0.001595095032826066, + 0.0014161107828840613, + 0.0015636079479008913, + 0.0011674945708364248, + 0.008052341639995575, + 0.0194729994982481, + 0.015566892921924591, + 0.03340604528784752, + 0.005583105143159628, + 0.005111855920404196, + 0.016097353771328926, + 0.006097207311540842, + 0.0018358981469646096, + 0.0023989405017346144, + 0.0018025117460638285, + 0.0023783876094967127, + 0.002076550154015422, + 0.0015708701685070992, + 0.001740285544656217, + 0.002065795473754406, + 0.002087112981826067, + 0.002624277723953128, + 0.0018325857818126678, + 0.0014387384289875627, + 0.0028689983300864697, + 0.0018588555976748466, + 0.0014981977874413133, + 0.0020405149552971125, + 0.0018215547315776348, + 0.0014239063020795584, + 0.0009913553949445486, + 0.000807581702247262, + 0.0015007219044491649, + 0.00117484200745821, + 0.0008171671652235091, + 0.0007770989905111492, + 0.0018921901937574148, + 0.0013905841624364257, + 0.0016877553425729275, + 0.0010290550999343395, + 0.000927356886677444, + 0.002595508238300681, + 0.0009070326341316104, + 0.002063787542283535, + 0.00290164933539927, + 0.0029674572870135307, + 0.002465247642248869, + 0.0022636267822235823, + 0.002421561861410737, + 0.0029399064369499683, + 0.00199608295224607, + 0.003200932638719678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0036353471223264933 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010480155469849706 + ], + "opposite_logprob_tokens": [ + -6.8760480880737305 + ], + "logprob_reward": -0.0010480155469849706, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010480155469849706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.556326687335968, + 0.015146640129387379, + 0.021555405110120773, + 0.017232252284884453, + 0.05474463105201721, + 0.06499675661325455, + 0.028166422620415688, + 0.024565594270825386, + 0.022540364414453506, + 0.19472530484199524 + ], + "credit_map": [ + 0.006907738745212555, + 0.015009572729468346, + 0.03567519783973694, + 0.25094878673553467, + 0.013284387066960335, + 0.0077962931245565414, + 0.21515141427516937, + 0.011553285643458366, + 0.0009690741426311433, + 0.0032540569081902504, + 0.0016473494470119476, + 0.001728324219584465, + 0.002684843959286809, + 0.001345754018984735, + 0.0016153142787516117, + 0.0019019230967387557, + 0.002652831142768264, + 0.002311919117346406, + 0.003307878505438566, + 0.0025711061898618937, + 0.002535693347454071, + 0.0023468323051929474, + 0.003480710554867983, + 0.0023484339471906424, + 0.0042446027509868145, + 0.0023631989024579525, + 0.002877169754356146, + 0.0019613008480519056, + 0.0012510762317106128, + 0.0012591041158884764, + 0.002109166933223605, + 0.0011666326317936182, + 0.003245366271585226, + 0.006474015302956104, + 0.006719175260514021, + 0.017927898094058037, + 0.004335759207606316, + 0.002987124491482973, + 0.009236513637006283, + 0.0038187766913324594, + 0.006514605600386858, + 0.010224191471934319, + 0.006274639628827572, + 0.008780152536928654, + 0.009124781005084515, + 0.0087733780965209, + 0.007211396005004644, + 0.008093615062534809, + 0.0033353003673255444, + 0.003804935608059168, + 0.0032168934121727943, + 0.0037766045425087214, + 0.004868418909609318, + 0.0028375161346048117, + 0.003042717929929495, + 0.0032840350177139044, + 0.0026274591218680143, + 0.0036083206068724394, + 0.0027463692240417004, + 0.0032674928661435843, + 0.00502848532050848, + 0.002364286221563816, + 0.0029091238975524902, + 0.002014053985476494, + 0.00216447445563972, + 0.0037472276017069817, + 0.0029537298250943422, + 0.0037216616328805685, + 0.0021395543590188026, + 0.0024226242676377296, + 0.0036037792451679707, + 0.0017873116303235292, + 0.027516458183526993, + 0.04241974279284477, + 0.024241259321570396, + 0.020621826872229576, + 0.03141575679183006, + 0.018534773960709572, + 0.01837322488427162, + 0.011602255515754223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0010480155469849706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349445804953575 + ], + "opposite_logprob_tokens": [ + -3.6513495445251465 + ], + "logprob_reward": -0.026349445804953575, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349445804953575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43248251080513, + 0.03639905899763107, + 0.09088167548179626, + 0.05767211690545082, + 0.08622287958860397, + 0.06580743938684464, + 0.037121567875146866, + 0.05135134980082512, + 0.07851303368806839, + 0.06354839354753494 + ], + "credit_map": [ + 0.008685625158250332, + 0.020222825929522514, + 0.025866448879241943, + 0.17991507053375244, + 0.018736744299530983, + 0.010187328793108463, + 0.15813474357128143, + 0.0107337087392807, + 0.004615535028278828, + 0.005670685786753893, + 0.005175387021154165, + 0.003374500200152397, + 0.0053338841535151005, + 0.004360722377896309, + 0.004231048747897148, + 0.003637293353676796, + 0.009048194624483585, + 0.010051892139017582, + 0.013838139362633228, + 0.015103654004633427, + 0.011084571480751038, + 0.008399770595133305, + 0.013247755356132984, + 0.010107689537107944, + 0.012444330379366875, + 0.009392456151545048, + 0.007839901372790337, + 0.006014890503138304, + 0.005657051224261522, + 0.005663844291120768, + 0.005435657687485218, + 0.005223982501775026, + 0.006334878504276276, + 0.013675987720489502, + 0.009868052788078785, + 0.024205666035413742, + 0.006602955516427755, + 0.0054078176617622375, + 0.013525513932108879, + 0.006602007430046797, + 0.006685993634164333, + 0.010822180658578873, + 0.006421716883778572, + 0.010937001556158066, + 0.006711878348141909, + 0.007824680767953396, + 0.008282925002276897, + 0.008121056482195854, + 0.0032600436825305223, + 0.004563839174807072, + 0.005609078798443079, + 0.007391514256596565, + 0.0037855005357414484, + 0.00336294062435627, + 0.005465869791805744, + 0.0036827789153903723, + 0.00813538022339344, + 0.008946684189140797, + 0.004817734006792307, + 0.004803855437785387, + 0.010482758283615112, + 0.00522806029766798, + 0.00425428943708539, + 0.004682580940425396, + 0.008182111196219921, + 0.011287489905953407, + 0.010152202099561691, + 0.014285686425864697, + 0.007419637870043516, + 0.008163834922015667, + 0.011635453440248966, + 0.007386614568531513, + 0.008008314296603203, + 0.013630497269332409, + 0.00754603510722518, + 0.005383524112403393, + 0.010470948182046413, + 0.007676207926124334, + 0.005757445935159922, + 0.005075416527688503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.026349445804953575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, the camera moving smoothly sideways, hugging the water's surface to keep pace with the subject. In the soft golden light of early morning, the vast, tranquil lake resembles a vast, unpolished silver mirror, reflecting the delicate colors of the sky and the distant, mist-shrouded mountain silhouettes. A swallow, its back feathers a deep blue-black sheen in the light, its belly a pristine, milky white, enters the frame with breathtaking speed. Its body is dramatically lowered, almost parallel to the water's surface, its streamlined form a picture of perfect aerodynamics. With powerful and graceful flapping of its long, pointed wings, each downward stroke, the tips of the wings slice into the water with exquisite precision and lightness, tandem, instantly carving brief, delicate V-shaped ripples across the glassy surface. These subtle ripples quickly spread outward, then swallowed by the lake's tranquility, leaving only a barely visible trail behind the swallow. The entire picture is full of dynamic beauty in tranquility, with soft light and shadow, and the tones are mainly cool lake blue and warm morning light, creating a poetic atmosphere that is both full of vitality and extremely peaceful and breathtaking.", + "seed": 16739, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_7/index_7621_seed_16739_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": -0.17447337079793215, + "vlm_logprob": -0.17447337079793215, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.17447337079793215, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5919446349143982 + ], + "opposite_logprob_tokens": [ + -0.9669446349143982 + ], + "logprob_reward": -0.5919446349143982, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5919446349143982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19863969087600708, + 0.14677779376506805, + 0.06665187329053879, + 0.03933526948094368, + 0.04222956299781799, + 0.07471415400505066, + 0.04989127814769745, + 0.04968086630105972, + 0.3068944811820984, + 0.025185024365782738 + ], + "credit_map": [ + 0.018927820026874542, + 0.043095700442790985, + 0.021729111671447754, + 0.026106808334589005, + 0.022769063711166382, + 0.01892545074224472, + 0.024306168779730797, + 0.02277955785393715, + 0.010887675918638706, + 0.041645362973213196, + 0.02180948108434677, + 0.014651532284915447, + 0.030126303434371948, + 0.008042634464800358, + 0.013616110198199749, + 0.0059986962005496025, + 0.009981597773730755, + 0.016406314447522163, + 0.00999350193887949, + 0.009172852151095867, + 0.006489811465144157, + 0.005841022823005915, + 0.0035489171277731657, + 0.005217862315475941, + 0.003923919051885605, + 0.005465167108923197, + 0.004988010041415691, + 0.006820205599069595, + 0.0039844391867518425, + 0.005508811678737402, + 0.00496270228177309, + 0.003682010807096958, + 0.0036800680682063103, + 0.0042725130915641785, + 0.009732237085700035, + 0.008215249516069889, + 0.004896797705441713, + 0.005167829804122448, + 0.0033060896676033735, + 0.0029587754979729652, + 0.009177548810839653, + 0.017622334882616997, + 0.01295542623847723, + 0.013189805671572685, + 0.006584485061466694, + 0.004374084062874317, + 0.003850099164992571, + 0.006960373837500811, + 0.005010024178773165, + 0.0062944223172962666, + 0.005397736094892025, + 0.006471308413892984, + 0.008207079023122787, + 0.0057276515290141106, + 0.006197056267410517, + 0.006586000788956881, + 0.0044095478951931, + 0.004504150245338678, + 0.006899931468069553, + 0.00953805074095726, + 0.00610900204628706, + 0.0066393520683050156, + 0.004946449771523476, + 0.006634386256337166, + 0.017362644895911217, + 0.016993967816233635, + 0.10466311126947403, + 0.020357735455036163, + 0.016924910247325897, + 0.01539719570428133, + 0.014793815091252327, + 0.10040108859539032, + 0.003147473558783531, + 0.003685588017106056, + 0.0033870076294988394, + 0.0032464517280459404, + 0.0036081853322684765, + 0.0024819059763103724, + 0.0023060012608766556, + 0.003322411561384797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.5919446349143982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22567006945610046 + ], + "opposite_logprob_tokens": [ + -1.6006700992584229 + ], + "logprob_reward": -0.22567006945610046, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22567006945610046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28992167115211487, + 0.07998305559158325, + 0.11793051660060883, + 0.055190011858940125, + 0.1131141185760498, + 0.04606524482369423, + 0.05319563299417496, + 0.04538818821310997, + 0.1608620434999466, + 0.03834952041506767 + ], + "credit_map": [ + 0.02560499496757984, + 0.04975225776433945, + 0.0331445150077343, + 0.04141249507665634, + 0.03029545582830906, + 0.02390913851559162, + 0.04692080244421959, + 0.038881998509168625, + 0.0065489597618579865, + 0.01715453714132309, + 0.014742961153388023, + 0.010898131877183914, + 0.012236488051712513, + 0.005783764645457268, + 0.00623389333486557, + 0.0063843210227787495, + 0.013382709585130215, + 0.02939668670296669, + 0.011956264264881611, + 0.012364023365080357, + 0.012577846646308899, + 0.014369022101163864, + 0.008555001579225063, + 0.015328963287174702, + 0.005496201105415821, + 0.010520550422370434, + 0.006974162068217993, + 0.007359934970736504, + 0.005760306026786566, + 0.0074646687135100365, + 0.006464334204792976, + 0.005149854812771082, + 0.008963726460933685, + 0.010421543382108212, + 0.02134370245039463, + 0.031463801860809326, + 0.012147115543484688, + 0.014238879084587097, + 0.008156117051839828, + 0.0063792322762310505, + 0.004314196761697531, + 0.007324635982513428, + 0.00725536048412323, + 0.005274528171867132, + 0.011013185605406761, + 0.0038375388830900192, + 0.0030502688605338335, + 0.003995527513325214, + 0.003899708855897188, + 0.006457297131419182, + 0.005667196121066809, + 0.008166223764419556, + 0.012254783883690834, + 0.006028157193213701, + 0.005320493597537279, + 0.0054017710499465466, + 0.004394565708935261, + 0.006530511658638716, + 0.0059668258763849735, + 0.007293920964002609, + 0.005134728271514177, + 0.006775300949811935, + 0.004330008756369352, + 0.004962327424436808, + 0.006919188424944878, + 0.009114854969084263, + 0.04703204333782196, + 0.01592198759317398, + 0.018554257228970528, + 0.00658218702301383, + 0.016148969531059265, + 0.04058855399489403, + 0.004276685882359743, + 0.00590337160974741, + 0.0054664150811731815, + 0.007010286208242178, + 0.004829473793506622, + 0.0025924723595380783, + 0.002703447360545397, + 0.005567367188632488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.22567006945610046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822710409760475 + ], + "opposite_logprob_tokens": [ + -4.0182271003723145 + ], + "logprob_reward": -0.01822710409760475, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822710409760475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05081122741103172, + 0.046672273427248, + 0.03589436039328575, + 0.07728037983179092, + 0.09415828436613083, + 0.042699020355939865, + 0.03817475587129593, + 0.06799328327178955, + 0.5246719717979431, + 0.021644454449415207 + ], + "credit_map": [ + 0.004388278350234032, + 0.009744607843458652, + 0.004983880091458559, + 0.005407844670116901, + 0.007927635684609413, + 0.004235656466335058, + 0.010062003508210182, + 0.004061317536979914, + 0.004432016983628273, + 0.012659856118261814, + 0.008310342207551003, + 0.005557440221309662, + 0.007064792327582836, + 0.0021617349702864885, + 0.004052747506648302, + 0.0024333447217941284, + 0.004088519141077995, + 0.008552414365112782, + 0.0036923708394169807, + 0.004907550290226936, + 0.006324581801891327, + 0.0029723020270466805, + 0.0021390854381024837, + 0.003217537421733141, + 0.007243688218295574, + 0.00913696363568306, + 0.007853349670767784, + 0.008596588857471943, + 0.010855461470782757, + 0.014669760130345821, + 0.013157769106328487, + 0.005766802467405796, + 0.00771030830219388, + 0.014067087322473526, + 0.018321609124541283, + 0.021853849291801453, + 0.011288907378911972, + 0.006519308313727379, + 0.007685041055083275, + 0.006712171249091625, + 0.0034981751814484596, + 0.005458890926092863, + 0.005855876486748457, + 0.005738730542361736, + 0.00892338901758194, + 0.005636652931571007, + 0.0035094302147626877, + 0.00407787412405014, + 0.0036326535046100616, + 0.004384817089885473, + 0.004549478646367788, + 0.004110863897949457, + 0.009278942830860615, + 0.00498588290065527, + 0.0035290648229420185, + 0.0037030524108558893, + 0.007513264659792185, + 0.01206181664019823, + 0.009117987006902695, + 0.009484225884079933, + 0.01277533546090126, + 0.005949874874204397, + 0.004674896597862244, + 0.006415884476155043, + 0.0224427692592144, + 0.028538012877106667, + 0.19327972829341888, + 0.03425008803606033, + 0.018320120871067047, + 0.018712233752012253, + 0.02779312990605831, + 0.18133586645126343, + 0.0025724112056195736, + 0.0027725277468562126, + 0.0031988543923944235, + 0.0021577670704573393, + 0.003752924269065261, + 0.002039640909060836, + 0.00244512059725821, + 0.002705208258703351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.01822710409760475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028257938101887703 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028257938101887703, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028257938101887703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06322695314884186, + 0.06056082621216774, + 0.028436411172151566, + 0.03170466423034668, + 0.059010960161685944, + 0.059501197189092636, + 0.032239437103271484, + 0.058284636586904526, + 0.5826557278633118, + 0.024379223585128784 + ], + "credit_map": [ + 0.004743895493447781, + 0.010552527382969856, + 0.006588305346667767, + 0.009217776358127594, + 0.00919469352811575, + 0.005176530219614506, + 0.010449660941958427, + 0.007303566671907902, + 0.00418228842318058, + 0.01259081531316042, + 0.009242679923772812, + 0.010973368771374226, + 0.009645935148000717, + 0.003357425332069397, + 0.0072300368919968605, + 0.0033382754772901535, + 0.003116422099992633, + 0.007068407256156206, + 0.0027310922741889954, + 0.0034427365753799677, + 0.005111356265842915, + 0.002260553417727351, + 0.002071246039122343, + 0.0026345946826040745, + 0.0034283031709492207, + 0.004939978010952473, + 0.0031473946291953325, + 0.003653095569461584, + 0.0035013086162507534, + 0.005881200544536114, + 0.0037447011563926935, + 0.0034086843952536583, + 0.005520177539438009, + 0.00990074872970581, + 0.010000866837799549, + 0.014445316046476364, + 0.005990070290863514, + 0.004645735491067171, + 0.004911466967314482, + 0.0035965803544968367, + 0.008061625063419342, + 0.015104603953659534, + 0.01141870953142643, + 0.007627436425536871, + 0.00573514960706234, + 0.003214925993233919, + 0.0026129386387765408, + 0.00572581123560667, + 0.00335855595767498, + 0.0035263807512819767, + 0.0037120275665074587, + 0.004136315081268549, + 0.006922951899468899, + 0.003789883805438876, + 0.0030792171601206064, + 0.0037141055800020695, + 0.006538379471749067, + 0.009701230563223362, + 0.007559291552752256, + 0.006975404918193817, + 0.010766339488327503, + 0.007419312838464975, + 0.0040952591225504875, + 0.005229420028626919, + 0.02349269390106201, + 0.02633976936340332, + 0.2006983458995819, + 0.04465879499912262, + 0.038879603147506714, + 0.024156350642442703, + 0.04153134673833847, + 0.18289878964424133, + 0.0021128030493855476, + 0.0028050176333636045, + 0.004127578344196081, + 0.002926427638158202, + 0.004336938261985779, + 0.002654217416420579, + 0.0022620249073952436, + 0.003154217032715678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0028257938101887703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 moving smoothly sideways while hugging the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a swallow visible with deep blue-black back feathers and a milky white belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the swallow fly with its body lowered almost parallel to the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the swallow's wing tips slice into the water creating brief V-shaped ripples?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369925171136856 + ], + "opposite_logprob_tokens": [ + -3.4086992740631104 + ], + "logprob_reward": -0.03369925171136856, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369925171136856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05527781322598457, + 0.15743082761764526, + 0.0468040369451046, + 0.013805892318487167, + 0.10022040456533432, + 0.0393996387720108, + 0.026770122349262238, + 0.03086358681321144, + 0.514528751373291, + 0.014898871071636677 + ], + "credit_map": [ + 0.0050714388489723206, + 0.012291046790778637, + 0.0049196332693099976, + 0.0071294368244707584, + 0.009922981262207031, + 0.004921041429042816, + 0.006477438844740391, + 0.004544796887785196, + 0.012829126790165901, + 0.04619000107049942, + 0.029650848358869553, + 0.012000052258372307, + 0.032209362834692, + 0.006506113335490227, + 0.013597079552710056, + 0.0044482396915555, + 0.0048516555689275265, + 0.013747661374509335, + 0.004911419469863176, + 0.004394969437271357, + 0.009525631554424763, + 0.0029790911357849836, + 0.0026223452296108007, + 0.0037712673656642437, + 0.0011921143159270287, + 0.0017111387569457293, + 0.001728906063362956, + 0.0023509806487709284, + 0.0012202406069263816, + 0.0024201984051615, + 0.0015721236122772098, + 0.0016101901419460773, + 0.010530918836593628, + 0.019001588225364685, + 0.019741838797926903, + 0.02574206329882145, + 0.007780731189996004, + 0.006070485804229975, + 0.006246947217732668, + 0.005105828400701284, + 0.004464510828256607, + 0.007726095616817474, + 0.007615106180310249, + 0.006130359601229429, + 0.004589603748172522, + 0.003144063986837864, + 0.002423160243779421, + 0.0033067367039620876, + 0.002058599377050996, + 0.0024813299532979727, + 0.003224357031285763, + 0.003977537155151367, + 0.005525155458599329, + 0.002914587501436472, + 0.0032516063656657934, + 0.0033369497396051884, + 0.002855054335668683, + 0.004964659921824932, + 0.004400521982461214, + 0.003127535106614232, + 0.005367507692426443, + 0.004676128271967173, + 0.0027141778264194727, + 0.0027580014429986477, + 0.019366180524230003, + 0.020884772762656212, + 0.18909354507923126, + 0.03468412160873413, + 0.02413354441523552, + 0.020099161192774773, + 0.030848197638988495, + 0.17541921138763428, + 0.001432007527910173, + 0.0021419026888906956, + 0.0018437851686030626, + 0.0024516519624739885, + 0.0021857316605746746, + 0.0014570876955986023, + 0.001166812377050519, + 0.0022198923397809267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.03369925171136856 + } + ], + "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/0000131/group_8/index_tensor([6151, 6151])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_8/index_tensor([6151, 6151])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7fc92b0304ad3d308f6b8c469546a8b29060dee1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_8/index_tensor([6151, 6151])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "perfectly steady, eye-level tracking shot places the sleek submersible perfectly at the center of the frame, gliding silently through the heartbreaking darkness of the abyssal depths. The vessel's hydrodynamic, manta-ray-like shape, constructed of seamless, polished obsidian alloy, reflects the scene's sole source of light: a vast, glowing tube winding through the dark, silty seafloor. This tube pulsates with an inner, futuristic energy, bathing the environment in alternating waves of vibrant cyan and electronic magenta light. The submersible moves effortlessly and steadily, its advanced propulsion system leaving no visible wake, only the faintest glimmer of light in the murky water. Intense neon lighting sharply defines the vessel's contours, painting one side a crisp cyan highlight and the other a deep burgundy sheen, while casting long, dramatic shadows that stretch into the infinite void. The overall atmosphere is one of breathtaking high-tech exploration, drenched in a futuristic neon aesthetic that contrasts the cold, lonely expanse of the deep sea with the vibrant pulse of advanced technology.", + "seed": 20827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_8/index_6151_seed_20827_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": -0.2643368514021859, + "vlm_logprob": -0.2643368514021859, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.2643368514021859, + "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 submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19074206054210663 + ], + "opposite_logprob_tokens": [ + -1.815742015838623 + ], + "logprob_reward": -0.19074206054210663, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19074206054210663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10234248638153076, + 0.14122362434864044, + 0.043380919843912125, + 0.05037551373243332, + 0.07140827924013138, + 0.0957629457116127, + 0.21233604848384857, + 0.06465350836515427, + 0.16086317598819733, + 0.05765355005860329 + ], + "credit_map": [ + 0.012055174447596073, + 0.01671072654426098, + 0.007873005233705044, + 0.012120668776333332, + 0.013799972832202911, + 0.011781251057982445, + 0.013046261854469776, + 0.014955426566302776, + 0.016692403703927994, + 0.011416214518249035, + 0.011437796987593174, + 0.017235945910215378, + 0.012429867871105671, + 0.03229083865880966, + 0.01501662377268076, + 0.02470392920076847, + 0.004336195532232523, + 0.0040734256617724895, + 0.003633933374658227, + 0.004074653144925833, + 0.005239543039351702, + 0.009352053515613079, + 0.002776600420475006, + 0.00989451352506876, + 0.009145736694335938, + 0.004205079283565283, + 0.005302774254232645, + 0.005666380748152733, + 0.0038946806453168392, + 0.009128380566835403, + 0.002793655963614583, + 0.010238823480904102, + 0.005096423905342817, + 0.005119432229548693, + 0.009556867182254791, + 0.008988242596387863, + 0.0038749997038394213, + 0.023100078105926514, + 0.0035754197742789984, + 0.012096814811229706, + 0.008830580860376358, + 0.011458217166364193, + 0.01605233922600746, + 0.01834803819656372, + 0.00878763385117054, + 0.015902118757367134, + 0.004672857467085123, + 0.011711152270436287, + 0.023517096415162086, + 0.012050521560013294, + 0.039093051105737686, + 0.03833121806383133, + 0.027573147788643837, + 0.029775982722640038, + 0.015430993400514126, + 0.026564037427306175, + 0.009128840640187263, + 0.005430137738585472, + 0.007540176622569561, + 0.008216657675802708, + 0.00528615852817893, + 0.009971468709409237, + 0.0032552529592067003, + 0.01582481525838375, + 0.02068975940346718, + 0.009006885811686516, + 0.015460457652807236, + 0.010102114640176296, + 0.014309457503259182, + 0.047137774527072906, + 0.009956119582057, + 0.03420060873031616, + 0.007323354948312044, + 0.004378502257168293, + 0.00761785451322794, + 0.00823291577398777, + 0.005536025855690241, + 0.010060898959636688, + 0.0047437832690775394, + 0.009760217741131783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.19074206054210663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006216049194336 + ], + "opposite_logprob_tokens": [ + -2.6950621604919434 + ], + "logprob_reward": -0.07006216049194336, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006216049194336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09178074449300766, + 0.08368145674467087, + 0.06494618207216263, + 0.0834774449467659, + 0.06560906022787094, + 0.19209659099578857, + 0.046720195561647415, + 0.095107302069664, + 0.1925530582666397, + 0.084027960896492 + ], + "credit_map": [ + 0.012192820198833942, + 0.010334371589124203, + 0.009024903178215027, + 0.012752169743180275, + 0.00832892395555973, + 0.015172146260738373, + 0.010390382260084152, + 0.013585032895207405, + 0.007328173611313105, + 0.0075635891407728195, + 0.004852988291531801, + 0.010804856196045876, + 0.011749038472771645, + 0.015686750411987305, + 0.011101123876869678, + 0.014594939537346363, + 0.003393730614334345, + 0.004164126235991716, + 0.0034807443153113127, + 0.00781667698174715, + 0.007383700460195541, + 0.0159293320029974, + 0.0055502778850495815, + 0.017227599397301674, + 0.0054137613624334335, + 0.0028208368457853794, + 0.003272938309237361, + 0.006226921919733286, + 0.002831487450748682, + 0.030633803457021713, + 0.002477450994774699, + 0.029800239950418472, + 0.004599851556122303, + 0.006459774449467659, + 0.005803137086331844, + 0.00613992428407073, + 0.003474077209830284, + 0.01744885928928852, + 0.003899296512827277, + 0.017784137278795242, + 0.023626601323485374, + 0.020380549132823944, + 0.04258481040596962, + 0.027268288657069206, + 0.012943437322974205, + 0.03603223338723183, + 0.009933046996593475, + 0.019327616319060326, + 0.005321687087416649, + 0.0035080257803201675, + 0.009044422768056393, + 0.008088753558695316, + 0.005924798082560301, + 0.0065963719971477985, + 0.00340962503105402, + 0.004826515447348356, + 0.011090458370745182, + 0.005495977587997913, + 0.013663941994309425, + 0.011667503044009209, + 0.006597471423447132, + 0.016181547194719315, + 0.005306129343807697, + 0.025104276835918427, + 0.02658011205494404, + 0.011151467449963093, + 0.017182286828756332, + 0.010585436597466469, + 0.01665261574089527, + 0.055377472192049026, + 0.011887038126587868, + 0.043136633932590485, + 0.014601676724851131, + 0.005671143531799316, + 0.008909069001674652, + 0.009808379225432873, + 0.00568007305264473, + 0.017855780199170113, + 0.00791033636778593, + 0.013591503724455833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.07006216049194336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948222242295742 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948222242295742, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948222242295742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1049136370420456, + 0.1183725968003273, + 0.056649476289749146, + 0.11964856833219528, + 0.0374276302754879, + 0.11087235808372498, + 0.12382198125123978, + 0.0853523388504982, + 0.06815364211797714, + 0.1747877299785614 + ], + "credit_map": [ + 0.021610630676150322, + 0.012357089668512344, + 0.012382355518639088, + 0.015603314153850079, + 0.0064389510080218315, + 0.01056844461709261, + 0.010296921245753765, + 0.01565592922270298, + 0.014418461360037327, + 0.009333380497992039, + 0.014012153260409832, + 0.016308650374412537, + 0.008585591800510883, + 0.02309974655508995, + 0.012611309066414833, + 0.02000330202281475, + 0.0032994698267430067, + 0.0016880222829058766, + 0.0039206272922456264, + 0.0039236885495483875, + 0.0022892290726304054, + 0.018847934901714325, + 0.0023418129421770573, + 0.020338693633675575, + 0.0058752428740262985, + 0.004562728106975555, + 0.005578896496444941, + 0.009331127628684044, + 0.0023346070665866137, + 0.04708268865942955, + 0.0023449817672371864, + 0.042538292706012726, + 0.0031066839583218098, + 0.002281203866004944, + 0.0038428897969424725, + 0.0034628750290721655, + 0.0021656136959791183, + 0.010697503574192524, + 0.002295837504789233, + 0.009575026109814644, + 0.014634594321250916, + 0.008399917744100094, + 0.024286892265081406, + 0.0188951063901186, + 0.00983351469039917, + 0.014976339414715767, + 0.0069751073606312275, + 0.012870877981185913, + 0.023225363343954086, + 0.009973645210266113, + 0.01568351872265339, + 0.012607867829501629, + 0.019407683983445168, + 0.01609417237341404, + 0.010162590071558952, + 0.016667133197188377, + 0.015799224376678467, + 0.004132697358727455, + 0.011369220912456512, + 0.008741438388824463, + 0.006119719240814447, + 0.012228290550410748, + 0.004645091947168112, + 0.02231665514409542, + 0.007504877634346485, + 0.0035743818152695894, + 0.004028677009046078, + 0.0030060913413763046, + 0.0035481336526572704, + 0.024430224671959877, + 0.0029573654755949974, + 0.01910388469696045, + 0.021826911717653275, + 0.011309999041259289, + 0.012990686111152172, + 0.016593551263213158, + 0.012142804451286793, + 0.05346021428704262, + 0.015617553144693375, + 0.030846014618873596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.005948222242295742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010411086259409785 + ], + "opposite_logprob_tokens": [ + -6.876040935516357 + ], + "logprob_reward": -0.0010411086259409785, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010411086259409785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11973445117473602, + 0.08420560508966446, + 0.1140972152352333, + 0.13594242930412292, + 0.05599929019808769, + 0.14548420906066895, + 0.18993814289569855, + 0.04546263441443443, + 0.06653983145952225, + 0.04259622097015381 + ], + "credit_map": [ + 0.018118299543857574, + 0.01608041301369667, + 0.01355675421655178, + 0.01682841219007969, + 0.010588891804218292, + 0.014628634788095951, + 0.011502268724143505, + 0.018430763855576515, + 0.009405313991010189, + 0.009996295906603336, + 0.007964159362018108, + 0.011729991994798183, + 0.006317747291177511, + 0.015680953860282898, + 0.008170339278876781, + 0.014940790832042694, + 0.0050901370123028755, + 0.0031124644447118044, + 0.0047567300498485565, + 0.0077978442423045635, + 0.002359459875151515, + 0.04202459752559662, + 0.003068093443289399, + 0.04588787257671356, + 0.005648607388138771, + 0.004586779512465, + 0.006811699364334345, + 0.009134029038250446, + 0.0038602815475314856, + 0.05205235630273819, + 0.0032706521451473236, + 0.05057801306247711, + 0.005718255415558815, + 0.005423239432275295, + 0.006405632011592388, + 0.00603674678131938, + 0.003415193408727646, + 0.012174061499536037, + 0.0035587800666689873, + 0.013267375528812408, + 0.015360532328486443, + 0.014737906865775585, + 0.029939670115709305, + 0.024043060839176178, + 0.009657781571149826, + 0.023374848067760468, + 0.007268845569342375, + 0.02110154740512371, + 0.027971824631094933, + 0.013480150140821934, + 0.030324216932058334, + 0.02061406522989273, + 0.027167219668626785, + 0.02480369806289673, + 0.017825454473495483, + 0.027751484885811806, + 0.007701634895056486, + 0.002946359571069479, + 0.0073413546197116375, + 0.007551902439445257, + 0.0031638306099921465, + 0.0065923151560127735, + 0.0029529002495110035, + 0.007212330587208271, + 0.013778292573988438, + 0.007177785504609346, + 0.007427325006574392, + 0.0058058323338627815, + 0.010047493502497673, + 0.008888804353773594, + 0.006788981147110462, + 0.006625307723879814, + 0.00396824162453413, + 0.0034386026673018932, + 0.003494709962978959, + 0.00418639974668622, + 0.003031620755791664, + 0.015325108543038368, + 0.0031280939001590014, + 0.006023438647389412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.0010411086259409785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0538907051086426 + ], + "opposite_logprob_tokens": [ + -0.4288907051086426 + ], + "logprob_reward": -1.0538907051086426, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538907051086426, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12318646907806396, + 0.08494766056537628, + 0.043088026344776154, + 0.08328516781330109, + 0.05370423197746277, + 0.11180723458528519, + 0.1322280615568161, + 0.06131445989012718, + 0.19189371168613434, + 0.11454492807388306 + ], + "credit_map": [ + 0.013846474699676037, + 0.015448677353560925, + 0.011902207508683205, + 0.018258461728692055, + 0.014069226570427418, + 0.012754089199006557, + 0.017895782366394997, + 0.019011564552783966, + 0.009596590884029865, + 0.007032050285488367, + 0.00670763710513711, + 0.01205860823392868, + 0.009376183152198792, + 0.015089473687112331, + 0.011727678589522839, + 0.013359441421926022, + 0.006244252435863018, + 0.004138090647757053, + 0.0047552501782774925, + 0.006047224160283804, + 0.0053779310546815395, + 0.005459560547024012, + 0.004145569633692503, + 0.006920150946825743, + 0.007352035026997328, + 0.003519094316288829, + 0.010667387396097183, + 0.007031856570392847, + 0.004370060283690691, + 0.023898666724562645, + 0.003294908907264471, + 0.023151163011789322, + 0.003932924475520849, + 0.0056069535203278065, + 0.005501181352883577, + 0.005507024936378002, + 0.002726909937337041, + 0.013000338338315487, + 0.003604210913181305, + 0.013824689202010632, + 0.012363431975245476, + 0.012368838302791119, + 0.018625685945153236, + 0.02202218770980835, + 0.009378403425216675, + 0.01775415614247322, + 0.005992020480334759, + 0.0133025161921978, + 0.018688803538680077, + 0.008732343092560768, + 0.024365970864892006, + 0.017044762149453163, + 0.016359416767954826, + 0.015693340450525284, + 0.014781830832362175, + 0.016561608761548996, + 0.010343261063098907, + 0.004874899983406067, + 0.009337021969258785, + 0.008685664273798466, + 0.005333594512194395, + 0.008868184871971607, + 0.004839992616325617, + 0.009031844325363636, + 0.02237080968916416, + 0.015279082581400871, + 0.017496302723884583, + 0.014208759181201458, + 0.023174472153186798, + 0.045668117702007294, + 0.019271697849035263, + 0.034424472600221634, + 0.010072970762848854, + 0.007769433781504631, + 0.009780999273061752, + 0.010145438835024834, + 0.009620086289942265, + 0.029355309903621674, + 0.011924367398023605, + 0.025876328349113464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -1.0538907051086426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "perfectly steady, eye-level tracking shot places the sleek submersible perfectly at the center of the frame, gliding silently through the heartbreaking darkness of the abyssal depths. The vessel's hydrodynamic, manta-ray-like shape, constructed of seamless, polished obsidian alloy, reflects the scene's sole source of light: a vast, glowing tube winding through the dark, silty seafloor. This tube pulsates with an inner, futuristic energy, bathing the environment in alternating waves of vibrant cyan and electronic magenta light. The submersible moves effortlessly and steadily, its advanced propulsion system leaving no visible wake, only the faintest glimmer of light in the murky water. Intense neon lighting sharply defines the vessel's contours, painting one side a crisp cyan highlight and the other a deep burgundy sheen, while casting long, dramatic shadows that stretch into the infinite void. The overall atmosphere is one of breathtaking high-tech exploration, drenched in a futuristic neon aesthetic that contrasts the cold, lonely expanse of the deep sea with the vibrant pulse of advanced technology.", + "seed": 20828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_8/index_6151_seed_20828_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": -0.35981077859178184, + "vlm_logprob": -0.35981077859178184, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.35981077859178184, + "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 submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7022326588630676 + ], + "opposite_logprob_tokens": [ + -0.7022326588630676 + ], + "logprob_reward": -0.7022326588630676, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7022326588630676, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24009335041046143, + 0.11392640322446823, + 0.03117993101477623, + 0.06365341693162918, + 0.038219988346099854, + 0.04835651069879532, + 0.09133611619472504, + 0.08663865178823471, + 0.2115945965051651, + 0.075001060962677 + ], + "credit_map": [ + 0.02253204584121704, + 0.020859621465206146, + 0.037557609379291534, + 0.032638344913721085, + 0.04600273445248604, + 0.0269891656935215, + 0.02551606297492981, + 0.027997763827443123, + 0.01121716108173132, + 0.009106006473302841, + 0.020885558798909187, + 0.024998987093567848, + 0.012547649443149567, + 0.012982659973204136, + 0.0104847876355052, + 0.011703585274517536, + 0.0016369608929380774, + 0.0023382985964417458, + 0.005331235472112894, + 0.006017754320055246, + 0.00515960156917572, + 0.0032129616010934114, + 0.003995202016085386, + 0.003487913403660059, + 0.0035120975226163864, + 0.003803180530667305, + 0.02115025743842125, + 0.012818999588489532, + 0.00628282455727458, + 0.00382611365057528, + 0.00902633648365736, + 0.0032336018048226833, + 0.003020291915163398, + 0.0032088360749185085, + 0.0060912203043699265, + 0.008540930226445198, + 0.0043457793071866035, + 0.0037669064477086067, + 0.003983824513852596, + 0.005262198857963085, + 0.0035435298923403025, + 0.006557341665029526, + 0.007754356600344181, + 0.008862541057169437, + 0.004266423173248768, + 0.006119364872574806, + 0.006091373041272163, + 0.00516157690435648, + 0.014330008998513222, + 0.0160971712321043, + 0.01451587863266468, + 0.013894730247557163, + 0.005368963349610567, + 0.009613349102437496, + 0.006169219966977835, + 0.011346787214279175, + 0.01030590571463108, + 0.010695191100239754, + 0.01117763388901949, + 0.007590922061353922, + 0.007847788743674755, + 0.014602587558329105, + 0.006771344691514969, + 0.01764727756381035, + 0.028142785653471947, + 0.028057295829057693, + 0.029728708788752556, + 0.017184175550937653, + 0.02211518958210945, + 0.024425402283668518, + 0.018833154812455177, + 0.04310787469148636, + 0.0072297086007893085, + 0.009838634170591831, + 0.00931065808981657, + 0.01118627842515707, + 0.008445890620350838, + 0.011796788312494755, + 0.009092514403164387, + 0.008100586012005806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.7022326588630676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297589540481567 + ], + "opposite_logprob_tokens": [ + -2.237975835800171 + ], + "logprob_reward": -0.11297589540481567, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297589540481567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13531732559204102, + 0.05803961679339409, + 0.08301866054534912, + 0.10791082680225372, + 0.06237058341503143, + 0.07644816488027573, + 0.028345346450805664, + 0.14833563566207886, + 0.12289401143789291, + 0.17731983959674835 + ], + "credit_map": [ + 0.013698853552341461, + 0.017191806808114052, + 0.023361532017588615, + 0.02020718716084957, + 0.01659589633345604, + 0.020238405093550682, + 0.010797800496220589, + 0.013225839473307133, + 0.0057202596217393875, + 0.005886916071176529, + 0.009326298721134663, + 0.013150440528988838, + 0.007088702637702227, + 0.006027934607118368, + 0.005548323504626751, + 0.005290740169584751, + 0.004614551085978746, + 0.007138811517506838, + 0.012789676897227764, + 0.010862544178962708, + 0.016268476843833923, + 0.010609537363052368, + 0.011991576291620731, + 0.00874349009245634, + 0.005303668323904276, + 0.00652239890769124, + 0.0437590517103672, + 0.020227864384651184, + 0.008820117451250553, + 0.003964661154896021, + 0.015250108204782009, + 0.004062955267727375, + 0.0043193562887609005, + 0.005996654275804758, + 0.009327576495707035, + 0.011924762278795242, + 0.006729185581207275, + 0.004195882938802242, + 0.010063746944069862, + 0.009813416749238968, + 0.006004902068525553, + 0.009100128896534443, + 0.021517721936106682, + 0.011229525320231915, + 0.004754415713250637, + 0.007041099946945906, + 0.007499193772673607, + 0.00930118653923273, + 0.004226814489811659, + 0.004357060883194208, + 0.004856305196881294, + 0.004110438749194145, + 0.0022213980555534363, + 0.0034772364888340235, + 0.001753143616952002, + 0.0033429472241550684, + 0.01458510011434555, + 0.01808059960603714, + 0.01877545937895775, + 0.01589609496295452, + 0.015235481783747673, + 0.027172178030014038, + 0.01570144109427929, + 0.022889282554388046, + 0.013499113731086254, + 0.01848234236240387, + 0.016327304765582085, + 0.010353493504226208, + 0.015416423790156841, + 0.020703673362731934, + 0.013147326186299324, + 0.014964338392019272, + 0.014811228029429913, + 0.018073445186018944, + 0.023588700219988823, + 0.026433389633893967, + 0.02261224389076233, + 0.02317800000309944, + 0.027930958196520805, + 0.020691869780421257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.11297589540481567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15654276311397552, + 0.166319340467453, + 0.04524511471390724, + 0.06407942622900009, + 0.07979071140289307, + 0.12924878299236298, + 0.04029949754476547, + 0.05905221775174141, + 0.19324487447738647, + 0.06617730110883713 + ], + "credit_map": [ + 0.01891600526869297, + 0.020975396037101746, + 0.024220991879701614, + 0.026530956849455833, + 0.0196754802018404, + 0.014485894702374935, + 0.014344335533678532, + 0.017393697053194046, + 0.014338490553200245, + 0.01502074021846056, + 0.02631254680454731, + 0.02894667163491249, + 0.024777451530098915, + 0.024984123185276985, + 0.019096653908491135, + 0.012842674739658833, + 0.0035886289551854134, + 0.004118265118449926, + 0.007153180427849293, + 0.007585315033793449, + 0.007863067090511322, + 0.005631597246974707, + 0.004630511626601219, + 0.0046745468862354755, + 0.003651101142168045, + 0.003387433011084795, + 0.020581722259521484, + 0.01748969964683056, + 0.0063155354000627995, + 0.0038903600070625544, + 0.00636757118627429, + 0.002396003343164921, + 0.0070877354592084885, + 0.01138340774923563, + 0.012655728496611118, + 0.019363591447472572, + 0.006179714109748602, + 0.00412830850109458, + 0.009261217899620533, + 0.009731007739901543, + 0.008489248342812061, + 0.014452734030783176, + 0.03742251172661781, + 0.018203234300017357, + 0.008004849776625633, + 0.010526375845074654, + 0.014700655825436115, + 0.01744917966425419, + 0.005148232448846102, + 0.007487741764634848, + 0.005978462751954794, + 0.0055937375873327255, + 0.002496544038876891, + 0.004322772845625877, + 0.002913961187005043, + 0.006358042825013399, + 0.007945084944367409, + 0.007416278123855591, + 0.009277709759771824, + 0.004816102329641581, + 0.006039315368980169, + 0.007700549438595772, + 0.005857277195900679, + 0.009999903850257397, + 0.027628937736153603, + 0.023456215858459473, + 0.027830902487039566, + 0.013549416325986385, + 0.01541425846517086, + 0.021962668746709824, + 0.013599413447082043, + 0.049803055822849274, + 0.007892352528870106, + 0.008982554078102112, + 0.007939374074339867, + 0.008888213895261288, + 0.006612118799239397, + 0.00967271625995636, + 0.006155068054795265, + 0.010034902021288872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.007630250882357359 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019405120983719826 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019405120983719826, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019405120983719826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20597843825817108, + 0.09755100309848785, + 0.04546383023262024, + 0.06173479184508324, + 0.08956421166658401, + 0.08448295295238495, + 0.05288955941796303, + 0.05604345351457596, + 0.21855640411376953, + 0.08773529529571533 + ], + "credit_map": [ + 0.017337890341877937, + 0.018621321767568588, + 0.03178204968571663, + 0.033226922154426575, + 0.04000542312860489, + 0.02137771062552929, + 0.020722612738609314, + 0.02290450595319271, + 0.009583084844052792, + 0.009991178289055824, + 0.018068071454763412, + 0.020502274855971336, + 0.008339222520589828, + 0.011922365985810757, + 0.008725712075829506, + 0.010419097729027271, + 0.0030381036922335625, + 0.005333473440259695, + 0.006827753968536854, + 0.008217991329729557, + 0.006773459259420633, + 0.004715026821941137, + 0.005649251863360405, + 0.0049087717197835445, + 0.00261209299787879, + 0.003620640141889453, + 0.01866784319281578, + 0.016105202957987785, + 0.004779008217155933, + 0.005017771385610104, + 0.007275544572621584, + 0.0036566865164786577, + 0.009367614053189754, + 0.009410311467945576, + 0.013320889323949814, + 0.017347540706396103, + 0.00900838989764452, + 0.006994516123086214, + 0.011518723331391811, + 0.01259621512144804, + 0.007986658252775669, + 0.015254136174917221, + 0.012140404433012009, + 0.010420156642794609, + 0.008340287022292614, + 0.010937124490737915, + 0.00978008285164833, + 0.009624099358916283, + 0.006676819641143084, + 0.00822629127651453, + 0.007769742514938116, + 0.006698754150420427, + 0.004167099483311176, + 0.008549414575099945, + 0.004220448434352875, + 0.006580994930118322, + 0.006774321664124727, + 0.008685708977282047, + 0.01147008128464222, + 0.0065331594087183475, + 0.004947577137500048, + 0.00658762501552701, + 0.003982629626989365, + 0.007062351796776056, + 0.027765795588493347, + 0.031010637059807777, + 0.029301783069968224, + 0.019619477912783623, + 0.02152823843061924, + 0.029291119426488876, + 0.020107636228203773, + 0.03993172571063042, + 0.009358036331832409, + 0.015160015784204006, + 0.010560878552496433, + 0.01146198995411396, + 0.007894887588918209, + 0.013874931260943413, + 0.007391271647065878, + 0.012033283710479736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0019405120983719826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.9742745757102966 + ], + "opposite_logprob_tokens": [ + -0.47427457571029663 + ], + "logprob_reward": -0.9742745757102966, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9742745757102966, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17611156404018402, + 0.15297597646713257, + 0.043321643024683, + 0.08979657292366028, + 0.04583923518657684, + 0.06103789433836937, + 0.05150057002902031, + 0.1092083677649498, + 0.17730647325515747, + 0.09290165454149246 + ], + "credit_map": [ + 0.02108711190521717, + 0.021935589611530304, + 0.02220067009329796, + 0.02473793551325798, + 0.02172245644032955, + 0.0242165420204401, + 0.02240239456295967, + 0.017808865755796432, + 0.013024114072322845, + 0.013684702105820179, + 0.024826133623719215, + 0.028362737968564034, + 0.017436973750591278, + 0.026883287355303764, + 0.015501949936151505, + 0.013256075792014599, + 0.0024369556922465563, + 0.0035760710015892982, + 0.006840379908680916, + 0.006609324365854263, + 0.006199313793331385, + 0.006604863330721855, + 0.005590797867625952, + 0.005463938694447279, + 0.0035570980980992317, + 0.003918815869837999, + 0.03358239307999611, + 0.02479318529367447, + 0.0069624618627130985, + 0.0030762231908738613, + 0.010829472914338112, + 0.0030769268050789833, + 0.004333410877734423, + 0.006723188329488039, + 0.006505614146590233, + 0.008474788628518581, + 0.0046236552298069, + 0.003022683784365654, + 0.006647791713476181, + 0.005508102010935545, + 0.006482656579464674, + 0.009615233168005943, + 0.01123564038425684, + 0.011942092329263687, + 0.003711892059072852, + 0.004824168514460325, + 0.005731660407036543, + 0.007494550198316574, + 0.009368269704282284, + 0.007858839817345142, + 0.006701415404677391, + 0.0069787753745913506, + 0.0034623732790350914, + 0.005551059264689684, + 0.003482813248410821, + 0.008097023703157902, + 0.01409329567104578, + 0.01469949446618557, + 0.018452893942594528, + 0.01609574630856514, + 0.008830521255731583, + 0.01313436683267355, + 0.008588945493102074, + 0.015313100069761276, + 0.01956893317401409, + 0.027927745133638382, + 0.0285014808177948, + 0.017748309299349785, + 0.01944567821919918, + 0.02249884605407715, + 0.016880428418517113, + 0.024735063314437866, + 0.008001413196325302, + 0.01716880314052105, + 0.01744616962969303, + 0.012508845888078213, + 0.00835834164172411, + 0.011572702787816525, + 0.0076576410792768, + 0.010187741369009018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.9742745757102966 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "perfectly steady, eye-level tracking shot places the sleek submersible perfectly at the center of the frame, gliding silently through the heartbreaking darkness of the abyssal depths. The vessel's hydrodynamic, manta-ray-like shape, constructed of seamless, polished obsidian alloy, reflects the scene's sole source of light: a vast, glowing tube winding through the dark, silty seafloor. This tube pulsates with an inner, futuristic energy, bathing the environment in alternating waves of vibrant cyan and electronic magenta light. The submersible moves effortlessly and steadily, its advanced propulsion system leaving no visible wake, only the faintest glimmer of light in the murky water. Intense neon lighting sharply defines the vessel's contours, painting one side a crisp cyan highlight and the other a deep burgundy sheen, while casting long, dramatic shadows that stretch into the infinite void. The overall atmosphere is one of breathtaking high-tech exploration, drenched in a futuristic neon aesthetic that contrasts the cold, lonely expanse of the deep sea with the vibrant pulse of advanced technology.", + "seed": 20829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_8/index_6151_seed_20829_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": -0.22900908635929226, + "vlm_logprob": -0.22900908635929226, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.22900908635929226, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26559433341026306 + ], + "opposite_logprob_tokens": [ + -1.5155943632125854 + ], + "logprob_reward": -0.26559433341026306, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26559433341026306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05714937672019005, + 0.06349051743745804, + 0.06888114660978317, + 0.04138968884944916, + 0.1008497104048729, + 0.14689363539218903, + 0.33209502696990967, + 0.11805569380521774, + 0.017768315970897675, + 0.05342691391706467 + ], + "credit_map": [ + 0.006876212079077959, + 0.006846167147159576, + 0.006971007212996483, + 0.008752009831368923, + 0.005739920772612095, + 0.005916743539273739, + 0.00884208083152771, + 0.007205238565802574, + 0.011133566498756409, + 0.00670134462416172, + 0.007918325252830982, + 0.006708244327455759, + 0.005778718739748001, + 0.008154754526913166, + 0.006493864580988884, + 0.010601697489619255, + 0.010015319101512432, + 0.0076356432400643826, + 0.008895312435925007, + 0.010360795073211193, + 0.007115203887224197, + 0.004665646236389875, + 0.011216615326702595, + 0.008976605720818043, + 0.00506290839985013, + 0.005105094518512487, + 0.006264700088649988, + 0.005683481227606535, + 0.004836405161768198, + 0.00362088019028306, + 0.006684083491563797, + 0.004132136236876249, + 0.01359595824033022, + 0.00912193488329649, + 0.015930194407701492, + 0.013371882028877735, + 0.007872113958001137, + 0.020942574366927147, + 0.0060238344594836235, + 0.013991217128932476, + 0.02116049826145172, + 0.013589496724307537, + 0.029876725748181343, + 0.02437572181224823, + 0.011845499277114868, + 0.014030593447387218, + 0.011587436310946941, + 0.020427672192454338, + 0.02697926200926304, + 0.029556473717093468, + 0.04858428239822388, + 0.04153342917561531, + 0.042100679129362106, + 0.055078521370887756, + 0.0374196320772171, + 0.050842754542827606, + 0.010952835902571678, + 0.012319032102823257, + 0.010359019972383976, + 0.008156955242156982, + 0.015801338478922844, + 0.027977673336863518, + 0.009550622664391994, + 0.022938210517168045, + 0.002153434557840228, + 0.0025653827469795942, + 0.0025353236123919487, + 0.0018242698861286044, + 0.0017598019912838936, + 0.0024980418384075165, + 0.002311208052560687, + 0.00212085316888988, + 0.006456606090068817, + 0.006047931499779224, + 0.005486282054334879, + 0.004859247710555792, + 0.005136095453053713, + 0.00681651197373867, + 0.010682687163352966, + 0.007941551506519318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.26559433341026306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016238272190094 + ], + "opposite_logprob_tokens": [ + -1.701623797416687 + ], + "logprob_reward": -0.2016238272190094, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016238272190094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10421840101480484, + 0.0701201856136322, + 0.06014319881796837, + 0.0724831372499466, + 0.08811122179031372, + 0.16325244307518005, + 0.12572665512561798, + 0.09492547810077667, + 0.06503073126077652, + 0.15598851442337036 + ], + "credit_map": [ + 0.012693449854850769, + 0.015418021939694881, + 0.01365758664906025, + 0.01595906727015972, + 0.00746299559250474, + 0.013864546082913876, + 0.011843213811516762, + 0.01331951841711998, + 0.010496113449335098, + 0.007111146580427885, + 0.008751270361244678, + 0.007836449891328812, + 0.007128216326236725, + 0.007607285398989916, + 0.00930539146065712, + 0.011884311214089394, + 0.007523848209530115, + 0.0059628733433783054, + 0.007254236843436956, + 0.007002301514148712, + 0.007712858263403177, + 0.005214113742113113, + 0.01100939605385065, + 0.008463571779429913, + 0.012596148997545242, + 0.006722199264913797, + 0.009363802149891853, + 0.00851866789162159, + 0.0093484902754426, + 0.008135350421071053, + 0.00921401847153902, + 0.008584459312260151, + 0.013081314042210579, + 0.0071532344445586205, + 0.014607781544327736, + 0.012891076505184174, + 0.006711184978485107, + 0.014094514772295952, + 0.007940733805298805, + 0.011631381697952747, + 0.024756379425525665, + 0.010348843410611153, + 0.029333917424082756, + 0.02232770062983036, + 0.0189177468419075, + 0.021860668435692787, + 0.012747878208756447, + 0.022959308698773384, + 0.022073715925216675, + 0.008275873027741909, + 0.019037948921322823, + 0.010925623588263988, + 0.009471338242292404, + 0.02269485779106617, + 0.01308179646730423, + 0.02016550488770008, + 0.009124180302023888, + 0.010977468453347683, + 0.009632647037506104, + 0.008407135494053364, + 0.01091891247779131, + 0.019447531551122665, + 0.01086499448865652, + 0.015552611090242863, + 0.008692660368978977, + 0.009191632270812988, + 0.007053384557366371, + 0.00905041303485632, + 0.00478840759024024, + 0.010573329403996468, + 0.006963027641177177, + 0.00871787779033184, + 0.019908586516976357, + 0.014323566108942032, + 0.019021926447749138, + 0.01666761189699173, + 0.014418506063520908, + 0.021006690338253975, + 0.02529963105916977, + 0.02534199319779873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.2016238272190094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306042939424515 + ], + "opposite_logprob_tokens": [ + -3.168060541152954 + ], + "logprob_reward": -0.04306042939424515, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306042939424515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1802942305803299, + 0.16172564029693604, + 0.17580316960811615, + 0.1616678237915039, + 0.09302700310945511, + 0.043860312551259995, + 0.052917759865522385, + 0.046759068965911865, + 0.060791872441768646, + 0.02315312996506691 + ], + "credit_map": [ + 0.02519119158387184, + 0.024221083149313927, + 0.02269003354012966, + 0.03317234292626381, + 0.010121959261596203, + 0.012786678969860077, + 0.02337847277522087, + 0.028732480481266975, + 0.03128339722752571, + 0.01818348839879036, + 0.021403245627880096, + 0.01910187117755413, + 0.009505818597972393, + 0.018097802996635437, + 0.017626537010073662, + 0.02652347832918167, + 0.020527269691228867, + 0.02139415219426155, + 0.02800004743039608, + 0.032169342041015625, + 0.010022570379078388, + 0.014860641211271286, + 0.028276439756155014, + 0.02055271528661251, + 0.03039737418293953, + 0.020440641790628433, + 0.020853787660598755, + 0.023803381249308586, + 0.009587484411895275, + 0.01643887348473072, + 0.020927613601088524, + 0.019218679517507553, + 0.010009588673710823, + 0.009249173104763031, + 0.01363404467701912, + 0.012716240249574184, + 0.012609169818460941, + 0.014401399530470371, + 0.008842513896524906, + 0.011564869433641434, + 0.006430936045944691, + 0.0042374408803880215, + 0.007191626820713282, + 0.006091439630836248, + 0.004669720306992531, + 0.005959125701338053, + 0.004738445859402418, + 0.004541582427918911, + 0.0052994112484157085, + 0.002626130124554038, + 0.007957009598612785, + 0.003419875167310238, + 0.002536908257752657, + 0.014495112001895905, + 0.004329313058406115, + 0.012253998778760433, + 0.005016759969294071, + 0.004321091342717409, + 0.004030210431665182, + 0.006928850896656513, + 0.004683975595980883, + 0.009736635722219944, + 0.0037806073669344187, + 0.00826093927025795, + 0.006622187327593565, + 0.00661895889788866, + 0.0063349464908242226, + 0.007503367029130459, + 0.005152545869350433, + 0.009660349227488041, + 0.008422894403338432, + 0.01047662552446127, + 0.0021006546448916197, + 0.002597304992377758, + 0.004280038643628359, + 0.0029274041298776865, + 0.0025622029788792133, + 0.00260895024985075, + 0.004115665797144175, + 0.001960909925401211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04306042939424515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019397982396185398 + ], + "opposite_logprob_tokens": [ + -6.25193977355957 + ], + "logprob_reward": -0.0019397982396185398, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019397982396185398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13787198066711426, + 0.050265707075595856, + 0.038037270307540894, + 0.0813884511590004, + 0.08497288078069687, + 0.08437559753656387, + 0.1774253100156784, + 0.1900475174188614, + 0.09030526131391525, + 0.06531000882387161 + ], + "credit_map": [ + 0.02063378505408764, + 0.017930252477526665, + 0.016079643741250038, + 0.01975778304040432, + 0.009328978136181831, + 0.01763479970395565, + 0.01905045099556446, + 0.017456280067563057, + 0.007843471132218838, + 0.005435447674244642, + 0.006484146695584059, + 0.0064318012446165085, + 0.00408794404938817, + 0.0058571407571434975, + 0.006731817033141851, + 0.007393933832645416, + 0.00560731440782547, + 0.004622842185199261, + 0.00523073086515069, + 0.004443123005330563, + 0.004959229379892349, + 0.003438587998971343, + 0.006059183273464441, + 0.0036762591917067766, + 0.008691680617630482, + 0.008839174173772335, + 0.012507675215601921, + 0.013456666842103004, + 0.0092374412342906, + 0.008771359920501709, + 0.012292865663766861, + 0.007591587956994772, + 0.011969326995313168, + 0.00952905509620905, + 0.012196771800518036, + 0.011533427983522415, + 0.01340440846979618, + 0.010169370099902153, + 0.006794154178351164, + 0.009376364760100842, + 0.007546728476881981, + 0.005622904747724533, + 0.01912650093436241, + 0.014811240136623383, + 0.0071152145974338055, + 0.008642545901238918, + 0.007683759555220604, + 0.013826706446707249, + 0.033354420214891434, + 0.016426587477326393, + 0.03347177803516388, + 0.016804594546556473, + 0.014798675663769245, + 0.026980699971318245, + 0.017093122005462646, + 0.018495438620448112, + 0.024210317060351372, + 0.01652419939637184, + 0.01938742958009243, + 0.0228318739682436, + 0.026268379762768745, + 0.03332981839776039, + 0.017579343169927597, + 0.029916150495409966, + 0.00841493159532547, + 0.013632744550704956, + 0.010139664635062218, + 0.014323027804493904, + 0.007316621486097574, + 0.014171337708830833, + 0.01093723252415657, + 0.01136969868093729, + 0.007828716188669205, + 0.005903730634599924, + 0.011837752535939217, + 0.008229500614106655, + 0.00579269602894783, + 0.009291489608585835, + 0.007073749788105488, + 0.009352367371320724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0019397982396185398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328270435333252 + ], + "opposite_logprob_tokens": [ + -0.7578270435333252 + ], + "logprob_reward": -0.6328270435333252, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328270435333252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16493014991283417, + 0.12650208175182343, + 0.04817463830113411, + 0.06745380908250809, + 0.05628550797700882, + 0.09515681117773056, + 0.18226580321788788, + 0.14160211384296417, + 0.06206795200705528, + 0.055561162531375885 + ], + "credit_map": [ + 0.021989034488797188, + 0.021361177787184715, + 0.0223500058054924, + 0.028406649827957153, + 0.008581060916185379, + 0.01700248010456562, + 0.01903880015015602, + 0.026200944557785988, + 0.020771002396941185, + 0.014858615584671497, + 0.01695231720805168, + 0.015667440369725227, + 0.00828552432358265, + 0.013703015632927418, + 0.014098369516432285, + 0.022165799513459206, + 0.007904977537691593, + 0.005162776913493872, + 0.006885104347020388, + 0.005201882217079401, + 0.004721452482044697, + 0.004184655379503965, + 0.008846856653690338, + 0.005266936030238867, + 0.014482568949460983, + 0.005617853719741106, + 0.008730738423764706, + 0.007796193473041058, + 0.006448714528232813, + 0.007469965610653162, + 0.006197470705956221, + 0.010710305534303188, + 0.00656007369980216, + 0.0055787572637200356, + 0.007199652027338743, + 0.010717394761741161, + 0.006781804375350475, + 0.007361512631177902, + 0.006732795853167772, + 0.005353518761694431, + 0.017986128106713295, + 0.008919496089220047, + 0.012876561842858791, + 0.015089811757206917, + 0.006867826450616121, + 0.0152759263291955, + 0.00913055706769228, + 0.00901049841195345, + 0.01194892916828394, + 0.010030612349510193, + 0.021690506488084793, + 0.016655081883072853, + 0.01346800196915865, + 0.05125654116272926, + 0.014540180563926697, + 0.042675938457250595, + 0.011708419770002365, + 0.01557181403040886, + 0.013088876381516457, + 0.009266123175621033, + 0.017478613182902336, + 0.03383537754416466, + 0.012042772956192493, + 0.0286101121455431, + 0.00763195613399148, + 0.0061972481198608875, + 0.007183348760008812, + 0.006136189680546522, + 0.008007479831576347, + 0.01007759664207697, + 0.0075899274088442326, + 0.009244207292795181, + 0.0063599576242268085, + 0.005662179086357355, + 0.008253649808466434, + 0.00514681451022625, + 0.005392885766923428, + 0.007499114144593477, + 0.01052603218704462, + 0.006720529869198799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.6328270435333252 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "perfectly steady, eye-level tracking shot places the sleek submersible perfectly at the center of the frame, gliding silently through the heartbreaking darkness of the abyssal depths. The vessel's hydrodynamic, manta-ray-like shape, constructed of seamless, polished obsidian alloy, reflects the scene's sole source of light: a vast, glowing tube winding through the dark, silty seafloor. This tube pulsates with an inner, futuristic energy, bathing the environment in alternating waves of vibrant cyan and electronic magenta light. The submersible moves effortlessly and steadily, its advanced propulsion system leaving no visible wake, only the faintest glimmer of light in the murky water. Intense neon lighting sharply defines the vessel's contours, painting one side a crisp cyan highlight and the other a deep burgundy sheen, while casting long, dramatic shadows that stretch into the infinite void. The overall atmosphere is one of breathtaking high-tech exploration, drenched in a futuristic neon aesthetic that contrasts the cold, lonely expanse of the deep sea with the vibrant pulse of advanced technology.", + "seed": 20830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_8/index_6151_seed_20830_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": -0.3015387696912512, + "vlm_logprob": -0.3015387696912512, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.3015387696912512, + "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 submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32493484020233154 + ], + "opposite_logprob_tokens": [ + -1.3249348402023315 + ], + "logprob_reward": -0.32493484020233154, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32493484020233154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08269015699625015, + 0.05882897973060608, + 0.021869562566280365, + 0.06256484985351562, + 0.0553918294608593, + 0.04322153702378273, + 0.1290278285741806, + 0.03838878497481346, + 0.41988250613212585, + 0.08813394606113434 + ], + "credit_map": [ + 0.00610375264659524, + 0.0123752411454916, + 0.009296929463744164, + 0.015567153692245483, + 0.007983645424246788, + 0.004886883310973644, + 0.012425603345036507, + 0.01405094563961029, + 0.003539108671247959, + 0.004006467759609222, + 0.011179150082170963, + 0.008761459030210972, + 0.003549321088939905, + 0.005032306537032127, + 0.008780347183346748, + 0.013980817049741745, + 0.0025099669583141804, + 0.0026320216711610556, + 0.0026530951727181673, + 0.002360240090638399, + 0.0031571045983582735, + 0.004129322711378336, + 0.0018111063400283456, + 0.002616703975945711, + 0.004796920344233513, + 0.008856138214468956, + 0.008830857463181019, + 0.010425711050629616, + 0.004704013001173735, + 0.008988328278064728, + 0.008117152377963066, + 0.007845732383430004, + 0.004949700552970171, + 0.005598984658718109, + 0.006371210794895887, + 0.012523162178695202, + 0.0054229432716965675, + 0.006618131883442402, + 0.006738869473338127, + 0.00716882711276412, + 0.004662064369767904, + 0.003989336080849171, + 0.006356023717671633, + 0.0056989118456840515, + 0.0036656572483479977, + 0.004287830553948879, + 0.00814727321267128, + 0.0064144376665353775, + 0.011332339607179165, + 0.04964694380760193, + 0.009600033052265644, + 0.005792501848191023, + 0.005434135440737009, + 0.0037643827963620424, + 0.0049325707368552685, + 0.03852493315935135, + 0.005207896698266268, + 0.005478092469274998, + 0.005322054494172335, + 0.0033340032678097486, + 0.005563008598983288, + 0.002926403423771262, + 0.004233940504491329, + 0.0063233887776732445, + 0.043603044003248215, + 0.06278327107429504, + 0.0655675157904625, + 0.04479442536830902, + 0.03527454659342766, + 0.0773434117436409, + 0.03911975398659706, + 0.051396530121564865, + 0.006559398025274277, + 0.009615590795874596, + 0.015574676916003227, + 0.021999459713697433, + 0.004435963463038206, + 0.006352977827191353, + 0.01304105669260025, + 0.010554825887084007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.32493484020233154 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298058182001114 + ], + "opposite_logprob_tokens": [ + -2.237980604171753 + ], + "logprob_reward": -0.11298058182001114, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298058182001114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09122131019830704, + 0.06339553743600845, + 0.049246083945035934, + 0.09398641437292099, + 0.053006403148174286, + 0.048288796097040176, + 0.3697513937950134, + 0.02324008010327816, + 0.12039414793252945, + 0.08746982365846634 + ], + "credit_map": [ + 0.006517291069030762, + 0.017408350482583046, + 0.009332386776804924, + 0.014487805776298046, + 0.012146961875259876, + 0.0072767180390655994, + 0.010059821419417858, + 0.013991974294185638, + 0.004804858472198248, + 0.003635380882769823, + 0.011646125465631485, + 0.009679648093879223, + 0.004740050528198481, + 0.005415049381554127, + 0.01098461914807558, + 0.0124898049980402, + 0.005475812125951052, + 0.006674257572740316, + 0.00840445514768362, + 0.005081034731119871, + 0.0058085317723453045, + 0.008092507719993591, + 0.0029266118071973324, + 0.006782869808375835, + 0.007179194130003452, + 0.013949016109108925, + 0.013051768764853477, + 0.016641149297356606, + 0.0067163631319999695, + 0.01254275068640709, + 0.011719566769897938, + 0.012186603620648384, + 0.007242664694786072, + 0.005004985257983208, + 0.005687238648533821, + 0.013599961064755917, + 0.004054551012814045, + 0.005629522260278463, + 0.005353391636162996, + 0.006434085313230753, + 0.0066659715957939625, + 0.005255969241261482, + 0.003907171543687582, + 0.004734506364911795, + 0.004974567797034979, + 0.0062937866896390915, + 0.007510931696742773, + 0.008945890702307224, + 0.02866416610777378, + 0.16021309792995453, + 0.011495593935251236, + 0.01127103716135025, + 0.016315793618559837, + 0.009105484001338482, + 0.00915108434855938, + 0.1235351711511612, + 0.002947855507954955, + 0.002634078962728381, + 0.0038111931644380093, + 0.002265722258016467, + 0.00293652038089931, + 0.0012468791101127863, + 0.002835171762853861, + 0.004562659654766321, + 0.017838766798377037, + 0.021509451791644096, + 0.01934943161904812, + 0.008346959948539734, + 0.008589928969740868, + 0.02151838131248951, + 0.00932393316179514, + 0.013917289674282074, + 0.00460510840639472, + 0.007957451045513153, + 0.018567411229014397, + 0.02610492892563343, + 0.0036213065031915903, + 0.003719389671459794, + 0.012676680460572243, + 0.010217546485364437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.11298058182001114 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192244037985802 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014192244037985802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192244037985802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1857432872056961, + 0.09488405287265778, + 0.051777433604002, + 0.14720000326633453, + 0.04314735159277916, + 0.04177907481789589, + 0.2953234910964966, + 0.023057959973812103, + 0.08400002866983414, + 0.03308728337287903 + ], + "credit_map": [ + 0.015676703304052353, + 0.029203632846474648, + 0.016970263794064522, + 0.02346905507147312, + 0.026418546214699745, + 0.03078867495059967, + 0.019611608237028122, + 0.02360481210052967, + 0.007143689319491386, + 0.008196162059903145, + 0.015147868543863297, + 0.011258049868047237, + 0.00944953877478838, + 0.008789094164967537, + 0.017347777262330055, + 0.01755187287926674, + 0.005675274413079023, + 0.005433478392660618, + 0.007559250108897686, + 0.008103815838694572, + 0.006428606808185577, + 0.00806402787566185, + 0.003878877032548189, + 0.006634104531258345, + 0.011832601390779018, + 0.011181640438735485, + 0.017981713637709618, + 0.025333954021334648, + 0.013664407655596733, + 0.02799411490559578, + 0.018000327050685883, + 0.021211257204413414, + 0.003835011273622513, + 0.0050731441006064415, + 0.00445192214101553, + 0.00947683397680521, + 0.006022620014846325, + 0.0048060305416584015, + 0.004563857335597277, + 0.004917931742966175, + 0.004326224327087402, + 0.003982254769653082, + 0.004104464780539274, + 0.005168549716472626, + 0.00388764007948339, + 0.006730949506163597, + 0.009582861326634884, + 0.0039961314760148525, + 0.025516605004668236, + 0.12278140336275101, + 0.0116130281239748, + 0.008918292820453644, + 0.012938864529132843, + 0.008721170015633106, + 0.00864725187420845, + 0.09618689864873886, + 0.003000120632350445, + 0.0033277792390435934, + 0.0030528828501701355, + 0.0026767277158796787, + 0.0025735197123140097, + 0.0017893100157380104, + 0.0027704723179340363, + 0.003867147956043482, + 0.01104781124740839, + 0.015048187226057053, + 0.013805532827973366, + 0.005015185102820396, + 0.0070059713907539845, + 0.015042919665575027, + 0.006544416770339012, + 0.01049000397324562, + 0.0030020338017493486, + 0.0030330971349030733, + 0.006087750196456909, + 0.00927751325070858, + 0.0010795985581353307, + 0.0013659741962328553, + 0.004171221517026424, + 0.005070093087852001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.014192244037985802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017124050064012408 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017124050064012408, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017124050064012408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026518072932958603, + 0.016225112602114677, + 0.021731873974204063, + 0.027266662567853928, + 0.025149589404463768, + 0.019671620801091194, + 0.7792086005210876, + 0.01706121303141117, + 0.055059660226106644, + 0.01210752036422491 + ], + "credit_map": [ + 0.002793649910017848, + 0.0034286919981241226, + 0.0036651701666414738, + 0.003993715159595013, + 0.0030335558112710714, + 0.003150199307128787, + 0.002552430145442486, + 0.003900660201907158, + 0.0013856234727427363, + 0.0019719013944268227, + 0.002033752389252186, + 0.0018442297587171197, + 0.001570056308992207, + 0.002448023296892643, + 0.0021418940741568804, + 0.002829632256180048, + 0.0017291391268372536, + 0.0018876264803111553, + 0.0028446947690099478, + 0.0025836338754743338, + 0.002950694877654314, + 0.005526477936655283, + 0.0015430356143042445, + 0.002666571643203497, + 0.0027128360234200954, + 0.004117449279874563, + 0.0033725693356245756, + 0.005158189684152603, + 0.0024503471795469522, + 0.0032581803388893604, + 0.002799214329570532, + 0.0033978745341300964, + 0.002571709221228957, + 0.003331416519358754, + 0.0035975126083940268, + 0.0033997013233602047, + 0.0029610099736601114, + 0.0034465522039681673, + 0.002933126874268055, + 0.002908561145886779, + 0.0022527670953422785, + 0.0023183354642242193, + 0.002100498415529728, + 0.0032849526032805443, + 0.0020748497918248177, + 0.002176160691305995, + 0.00308221485465765, + 0.0023818407207727432, + 0.06256167590618134, + 0.33867892622947693, + 0.02147333323955536, + 0.022879133000969887, + 0.03372817859053612, + 0.01983058452606201, + 0.019252177327871323, + 0.2608046233654022, + 0.0021565877832472324, + 0.0027307062409818172, + 0.002325764624401927, + 0.0016557590570300817, + 0.0020602820441126823, + 0.0010780917946249247, + 0.0018321850802749395, + 0.003221836406737566, + 0.004367790184915066, + 0.006436533760279417, + 0.007388299331068993, + 0.00912444293498993, + 0.004305836744606495, + 0.009372562170028687, + 0.007286583539098501, + 0.006777611561119556, + 0.0009845609311014414, + 0.00105886603705585, + 0.002452600048854947, + 0.002450787229463458, + 0.0003766916925087571, + 0.0007290635840035975, + 0.0013797138817608356, + 0.0026752364356070757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.0017124050064012408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the submersible positioned at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the submersible constructed of a seamless, polished obsidian alloy with a manta-ray-like shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, silty seafloor featuring a vast glowing tube winding through it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense neon lighting with alternating waves of vibrant cyan and electronic magenta?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot perfectly steady and eye-level while tracking the submersible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0538737773895264 + ], + "opposite_logprob_tokens": [ + -0.42887380719184875 + ], + "logprob_reward": -1.0538737773895264, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538737773895264, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05687526986002922, + 0.034004855901002884, + 0.07739454507827759, + 0.037098973989486694, + 0.046037036925554276, + 0.05523262917995453, + 0.42962440848350525, + 0.03239060938358307, + 0.20007726550102234, + 0.03126446530222893 + ], + "credit_map": [ + 0.005316175054758787, + 0.009820577688515186, + 0.007013126276433468, + 0.00749434856697917, + 0.006394216790795326, + 0.004342011176049709, + 0.0055026765912771225, + 0.010992134921252728, + 0.003003505291417241, + 0.0023307213559746742, + 0.005859297700226307, + 0.0036664986982941628, + 0.002545024733990431, + 0.0029366789385676384, + 0.006146211642771959, + 0.007516918703913689, + 0.010678429156541824, + 0.008017833344638348, + 0.010633762925863266, + 0.00792375486344099, + 0.008187182247638702, + 0.016418108716607094, + 0.005758693907409906, + 0.00977677945047617, + 0.0036081636790186167, + 0.002270642900839448, + 0.0038170358166098595, + 0.004327220376580954, + 0.005003228783607483, + 0.008939582854509354, + 0.0045526172034442425, + 0.004580484237521887, + 0.004424169659614563, + 0.00584468524903059, + 0.0057133762165904045, + 0.008835041895508766, + 0.006134560331702232, + 0.004621776286512613, + 0.005320081487298012, + 0.005143347661942244, + 0.006731573026627302, + 0.0057675037533044815, + 0.0049110217951238155, + 0.008831292390823364, + 0.004094312898814678, + 0.006147786509245634, + 0.007062832824885845, + 0.011686304584145546, + 0.03493963181972504, + 0.18511247634887695, + 0.013012250885367393, + 0.013560733757913113, + 0.018931232392787933, + 0.010929180309176445, + 0.011670813895761967, + 0.1414680927991867, + 0.0035298275761306286, + 0.003699769265949726, + 0.004619141109287739, + 0.0023033618927001953, + 0.006755738984793425, + 0.002530846744775772, + 0.004326818976551294, + 0.004625104833394289, + 0.01941632106900215, + 0.028296301141381264, + 0.03161494433879852, + 0.02546936459839344, + 0.01648639887571335, + 0.03437206149101257, + 0.01939498633146286, + 0.025026880204677582, + 0.003204021602869034, + 0.003301463322713971, + 0.005083888303488493, + 0.0065463269129395485, + 0.0022508902475237846, + 0.002258294029161334, + 0.0042408667504787445, + 0.004378712736070156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -1.0538737773895264 + } + ], + "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/0000131/group_9/index_tensor([3867, 3867])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_9/index_tensor([3867, 3867])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..797db14b1b0c70a6d95abe3c678550804334fd29 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_9/index_tensor([3867, 3867])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, slowly and steadily advancing, focuses on a pale forearm of a figure, set against a soft, out-of-focus background. Soft, diffused morning light from an unseen source illuminates the scene, capturing the tiny, almost invisible hairs on the skin. A delicate network of shallow, healing wounds dot the forearm, their pale pink lines contrasting sharply with the surrounding skin. Miraculously, from these wounds, vibrant green sprouts begin to carefully break through the surface, not with force but with gentle, organic persistence. The camera's slow advance allows us to witness these incredibly fragile sprouts steadily elongate and unfold tiny, perfectly shaped leaves that shimmer as if kissed by dewdrops. The overall atmosphere is one of quiet, surreal wonder, blending the visceral reality of scarred skin with the quiet, magical beauty of new life growing through the passage of time, in a continuous and graceful movement.", + "seed": 4862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_9/index_3867_seed_4862_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": -0.01235029436647892, + "vlm_logprob": -0.01235029436647892, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.01235029436647892, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.003822521772235632 + ], + "opposite_logprob_tokens": [ + -7.378822326660156 + ], + "logprob_reward": -0.003822521772235632, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.003822521772235632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25776439905166626, + 0.04549489542841911, + 0.04995109513401985, + 0.17034536600112915, + 0.14604954421520233, + 0.04500425606966019, + 0.14118623733520508, + 0.058843303471803665, + 0.03559150546789169, + 0.04976940527558327 + ], + "credit_map": [ + 0.014560393989086151, + 0.02151278778910637, + 0.026198124513030052, + 0.04381529241800308, + 0.03583801910281181, + 0.04917309805750847, + 0.0376652367413044, + 0.029001450166106224, + 0.0027548708021640778, + 0.0031819327268749475, + 0.0028805406764149666, + 0.005978174041956663, + 0.00500086322426796, + 0.012209564447402954, + 0.0075907399877905846, + 0.005898207891732454, + 0.0034353246446698904, + 0.004558323882520199, + 0.009248999878764153, + 0.007154352962970734, + 0.006092946510761976, + 0.007455745711922646, + 0.007419016677886248, + 0.004586384166032076, + 0.009218567982316017, + 0.009503394365310669, + 0.01461161207407713, + 0.04701288044452667, + 0.018447130918502808, + 0.03088158182799816, + 0.023876363411545753, + 0.016793852671980858, + 0.010814385488629341, + 0.009780608117580414, + 0.010212930850684643, + 0.037309613078832626, + 0.020071018487215042, + 0.02534843049943447, + 0.01704835332930088, + 0.015464212745428085, + 0.003497185418382287, + 0.005766555666923523, + 0.003986767027527094, + 0.005221097264438868, + 0.004407969303429127, + 0.009638825431466103, + 0.004630803130567074, + 0.007855058647692204, + 0.007266481872648001, + 0.009176109917461872, + 0.006463348865509033, + 0.009756858460605145, + 0.008783027529716492, + 0.04730738699436188, + 0.010344334878027439, + 0.042088691145181656, + 0.002851151628419757, + 0.002313113072887063, + 0.004127283580601215, + 0.005938567221164703, + 0.0032527227886021137, + 0.01888336054980755, + 0.004422794561833143, + 0.01705431006848812, + 0.002723835175856948, + 0.004656607750803232, + 0.0029136643279343843, + 0.00394434854388237, + 0.004037074279040098, + 0.007201635278761387, + 0.0037869035732001066, + 0.006327443290501833, + 0.004432962276041508, + 0.003816836979240179, + 0.0033269133418798447, + 0.005807643756270409, + 0.006625725422054529, + 0.01151441689580679, + 0.005193450953811407, + 0.009051456116139889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.003822521772235632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.005953317973762751 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953317973762751, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953317973762751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10214180499315262, + 0.16985434293746948, + 0.09811260551214218, + 0.05796942487359047, + 0.22557178139686584, + 0.03218187391757965, + 0.17372658848762512, + 0.06525922566652298, + 0.043769508600234985, + 0.031412869691848755 + ], + "credit_map": [ + 0.006675302982330322, + 0.010981138795614243, + 0.010543476790189743, + 0.01834576576948166, + 0.015999726951122284, + 0.017699439078569412, + 0.01438289787620306, + 0.007514051627367735, + 0.008300121873617172, + 0.012201910838484764, + 0.011949503794312477, + 0.02975546382367611, + 0.03089054301381111, + 0.028152046725153923, + 0.029578199610114098, + 0.01902652718126774, + 0.0051489826291799545, + 0.006191245745867491, + 0.007268485613167286, + 0.019114134833216667, + 0.01213753130286932, + 0.01991424523293972, + 0.015566068701446056, + 0.012771903537213802, + 0.003316243179142475, + 0.002775677712634206, + 0.0043276152573525906, + 0.013795214705169201, + 0.009544567205011845, + 0.00943593680858612, + 0.009682303294539452, + 0.005091861821711063, + 0.015269581228494644, + 0.008996779099106789, + 0.014211341738700867, + 0.06415244936943054, + 0.02852325513958931, + 0.043771181255578995, + 0.02772604301571846, + 0.02292114496231079, + 0.002088117180392146, + 0.004070827271789312, + 0.002967579523101449, + 0.0055456096306443214, + 0.004659709054976702, + 0.004666063003242016, + 0.004798452369868755, + 0.0033855156507343054, + 0.008443621918559074, + 0.010155607014894485, + 0.006725668907165527, + 0.011396355926990509, + 0.008698997087776661, + 0.060669444501399994, + 0.01264278031885624, + 0.05499409884214401, + 0.0046045719645917416, + 0.003700317582115531, + 0.0045130206272006035, + 0.0065851109102368355, + 0.004090539645403624, + 0.018600326031446457, + 0.006042585242539644, + 0.017122741788625717, + 0.0026499531231820583, + 0.0030884782318025827, + 0.0030748688150197268, + 0.00434380117803812, + 0.003823487786576152, + 0.010870598256587982, + 0.007682605180889368, + 0.008235708810389042, + 0.0022972740698605776, + 0.008443055674433708, + 0.0035313712432980537, + 0.002935332478955388, + 0.003775551449507475, + 0.004464091267436743, + 0.003083212533965707, + 0.0028829763177782297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.005953317973762751 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.004104286432266235 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004104286432266235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104286432266235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20085018873214722, + 0.10698062926530838, + 0.054547376930713654, + 0.07927636057138443, + 0.22991591691970825, + 0.04465582221746445, + 0.04873307794332504, + 0.0706334188580513, + 0.11102312803268433, + 0.05338403955101967 + ], + "credit_map": [ + 0.011667265556752682, + 0.020923689007759094, + 0.023683954030275345, + 0.03486394137144089, + 0.02617429569363594, + 0.029813658446073532, + 0.02846568077802658, + 0.02525772713124752, + 0.004516612272709608, + 0.007301441393792629, + 0.012802720069885254, + 0.023187750950455666, + 0.012809522449970245, + 0.019805990159511566, + 0.015544059686362743, + 0.011012538336217403, + 0.0040752943605184555, + 0.005217316094785929, + 0.006044303998351097, + 0.0066100312396883965, + 0.0066785323433578014, + 0.010254506953060627, + 0.009184541180729866, + 0.006482857745140791, + 0.003975620958954096, + 0.0035034839529544115, + 0.0061550927348434925, + 0.021777037531137466, + 0.010382503271102905, + 0.013509129174053669, + 0.012467135675251484, + 0.007506364956498146, + 0.014805139042437077, + 0.00925314612686634, + 0.014536884613335133, + 0.06749255955219269, + 0.028317559510469437, + 0.044244203716516495, + 0.02761813998222351, + 0.023648323491215706, + 0.0034599110949784517, + 0.006081435363739729, + 0.004014086443930864, + 0.006243772804737091, + 0.0054467180743813515, + 0.0067659299820661545, + 0.005965915042907, + 0.00667805690318346, + 0.0033832332119345665, + 0.004040594212710857, + 0.0027794199995696545, + 0.004627867601811886, + 0.005096327513456345, + 0.01216752640902996, + 0.005646948702633381, + 0.010991166345775127, + 0.005785221699625254, + 0.006221027113497257, + 0.005988576915115118, + 0.00904246885329485, + 0.007692803628742695, + 0.013747773133218288, + 0.009405291639268398, + 0.012750254943966866, + 0.005958652123808861, + 0.00758713111281395, + 0.00514157023280859, + 0.0072972895577549934, + 0.007949348539113998, + 0.035873591899871826, + 0.011044972576200962, + 0.030170582234859467, + 0.004380695056170225, + 0.009744906798005104, + 0.005684096831828356, + 0.005513735581189394, + 0.006243651267141104, + 0.007375645451247692, + 0.008242099545896053, + 0.006199211347848177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.004104286432266235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.038090646266937256 + ], + "opposite_logprob_tokens": [ + -3.288090705871582 + ], + "logprob_reward": -0.038090646266937256, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038090646266937256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11835066229104996, + 0.06383579969406128, + 0.039306897670030594, + 0.20050854980945587, + 0.14625796675682068, + 0.038323014974594116, + 0.1840287148952484, + 0.0570940226316452, + 0.03977326303720474, + 0.11252117156982422 + ], + "credit_map": [ + 0.0058712647296488285, + 0.009759895503520966, + 0.01218656450510025, + 0.019517647102475166, + 0.01668071560561657, + 0.025938738137483597, + 0.018884262070059776, + 0.009511556476354599, + 0.0029650956857949495, + 0.004604907240718603, + 0.005900754127651453, + 0.014713418669998646, + 0.007189975120127201, + 0.013986261561512947, + 0.00893324613571167, + 0.005542132072150707, + 0.0031532400753349066, + 0.0026440632063895464, + 0.0037472054827958345, + 0.005682486575096846, + 0.004645006265491247, + 0.007815019227564335, + 0.005450544413179159, + 0.006169328931719065, + 0.011091266758739948, + 0.01118929497897625, + 0.017875853925943375, + 0.05431997403502464, + 0.022367965430021286, + 0.035009969025850296, + 0.028692642226815224, + 0.019961565732955933, + 0.009621446952223778, + 0.009469087235629559, + 0.010239394381642342, + 0.04056767374277115, + 0.01916046440601349, + 0.02458363212645054, + 0.017056329175829887, + 0.015559918247163296, + 0.002802015747874975, + 0.005512382369488478, + 0.0037088620010763407, + 0.007339373230934143, + 0.004164988175034523, + 0.005476083606481552, + 0.005410476122051477, + 0.003908827900886536, + 0.008162970654666424, + 0.01057278923690319, + 0.007591908797621727, + 0.01143121998757124, + 0.009740770794451237, + 0.06498488783836365, + 0.013388856314122677, + 0.05815529450774193, + 0.0030038179829716682, + 0.0038106057327240705, + 0.002287522191181779, + 0.005492204334586859, + 0.0034572759177535772, + 0.01666303351521492, + 0.006018457468599081, + 0.016361098736524582, + 0.0026279850862920284, + 0.005119824782013893, + 0.003040363546460867, + 0.00557438749819994, + 0.004844526294618845, + 0.005846809595823288, + 0.006231518927961588, + 0.006487841717898846, + 0.007879313081502914, + 0.011213679797947407, + 0.011564142070710659, + 0.012140138074755669, + 0.010419806465506554, + 0.020425153896212578, + 0.017689911648631096, + 0.021189015358686447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.038090646266937256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.009780699387192726 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780699387192726, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780699387192726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06915431469678879, + 0.12638187408447266, + 0.08513280749320984, + 0.07924863696098328, + 0.26401108503341675, + 0.04517016187310219, + 0.07381582260131836, + 0.0739813819527626, + 0.04249038174748421, + 0.14061354100704193 + ], + "credit_map": [ + 0.00476064020767808, + 0.007250321563333273, + 0.0078041888773441315, + 0.011350751854479313, + 0.010437151417136192, + 0.011510826647281647, + 0.009396490640938282, + 0.0066439430229365826, + 0.005642311181873083, + 0.008019913919270039, + 0.01297011598944664, + 0.028459956869482994, + 0.013499035499989986, + 0.030247140675783157, + 0.018990397453308105, + 0.008553006686270237, + 0.006021539214998484, + 0.011012914590537548, + 0.013685339130461216, + 0.011358028277754784, + 0.010096634738147259, + 0.012287919409573078, + 0.011393630877137184, + 0.009276802651584148, + 0.004888053983449936, + 0.005136718042194843, + 0.004924636799842119, + 0.02002701535820961, + 0.009204390458762646, + 0.014632373116910458, + 0.011855661869049072, + 0.008579784072935581, + 0.01837817020714283, + 0.011669686995446682, + 0.01757901720702648, + 0.07332215458154678, + 0.029689183458685875, + 0.05403662100434303, + 0.03266754746437073, + 0.026668693870306015, + 0.0038679398130625486, + 0.0074465773068368435, + 0.0042415461502969265, + 0.006464339792728424, + 0.005640310700982809, + 0.00740409828722477, + 0.004459771327674389, + 0.005645575933158398, + 0.0031986695248633623, + 0.005117164924740791, + 0.0033509708009660244, + 0.005145757459104061, + 0.004228179808706045, + 0.024274280294775963, + 0.005655167158693075, + 0.02284562960267067, + 0.005959776695817709, + 0.0025937373284250498, + 0.005257439333945513, + 0.007237826008349657, + 0.0031672653276473284, + 0.023428495973348618, + 0.006058060564100742, + 0.02027878351509571, + 0.0023707475047558546, + 0.0035462325904518366, + 0.0024996339343488216, + 0.0030623129568994045, + 0.002219976158812642, + 0.01277481485158205, + 0.0036178415175527334, + 0.01239881943911314, + 0.01213214173913002, + 0.0185715164989233, + 0.0232651736587286, + 0.014064395800232887, + 0.014011639170348644, + 0.020953809842467308, + 0.01541250478476286, + 0.02220236510038376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.009780699387192726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, slowly and steadily advancing, focuses on a pale forearm of a figure, set against a soft, out-of-focus background. Soft, diffused morning light from an unseen source illuminates the scene, capturing the tiny, almost invisible hairs on the skin. A delicate network of shallow, healing wounds dot the forearm, their pale pink lines contrasting sharply with the surrounding skin. Miraculously, from these wounds, vibrant green sprouts begin to carefully break through the surface, not with force but with gentle, organic persistence. The camera's slow advance allows us to witness these incredibly fragile sprouts steadily elongate and unfold tiny, perfectly shaped leaves that shimmer as if kissed by dewdrops. The overall atmosphere is one of quiet, surreal wonder, blending the visceral reality of scarred skin with the quiet, magical beauty of new life growing through the passage of time, in a continuous and graceful movement.", + "seed": 4863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_9/index_3867_seed_4863_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": -0.012093111616559326, + "vlm_logprob": -0.012093111616559326, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.012093111616559326, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.0016337393317371607 + ], + "opposite_logprob_tokens": [ + -7.251633644104004 + ], + "logprob_reward": -0.0016337393317371607, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0016337393317371607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250409632921219, + 0.03108256310224533, + 0.02289414033293724, + 0.024177709594368935, + 0.048980750143527985, + 0.027787089347839355, + 0.0893017053604126, + 0.027643758803606033, + 0.44164490699768066, + 0.16144640743732452 + ], + "credit_map": [ + 0.008412241004407406, + 0.009737619198858738, + 0.026440903544425964, + 0.019373713061213493, + 0.014826246537268162, + 0.007787124253809452, + 0.008888788521289825, + 0.029574330896139145, + 0.003555804258212447, + 0.003261665115132928, + 0.004495155066251755, + 0.004802072886377573, + 0.005763818044215441, + 0.002635418204590678, + 0.003323894226923585, + 0.0032447362318634987, + 0.003374980529770255, + 0.0018965033814311028, + 0.002323808614164591, + 0.004282734822481871, + 0.003281578654423356, + 0.001980922883376479, + 0.002375308657065034, + 0.0033783037215471268, + 0.0030857280362397432, + 0.001643796218559146, + 0.003504713298752904, + 0.003725627204403281, + 0.0036829914897680283, + 0.004109140019863844, + 0.0026149095501750708, + 0.0018108010990545154, + 0.006010302342474461, + 0.00452113151550293, + 0.014337620697915554, + 0.0031795979011803865, + 0.002910909941419959, + 0.003898013150319457, + 0.0033437928650528193, + 0.010779382660984993, + 0.0032632481306791306, + 0.0027168833184987307, + 0.004264365416020155, + 0.004866821691393852, + 0.003304651938378811, + 0.0042738826014101505, + 0.0026751423720270395, + 0.002422094577923417, + 0.01692076399922371, + 0.00416838051751256, + 0.033707667142152786, + 0.005661524832248688, + 0.005863246042281389, + 0.007584851235151291, + 0.007909553125500679, + 0.007485717069357634, + 0.005372989922761917, + 0.002918432466685772, + 0.004013786558061838, + 0.002945195185020566, + 0.0025420025922358036, + 0.0030957909766584635, + 0.003403655719012022, + 0.003351903520524502, + 0.057442065328359604, + 0.02639162726700306, + 0.08868546038866043, + 0.036283235996961594, + 0.02872166782617569, + 0.07866919040679932, + 0.043120820075273514, + 0.0823308527469635, + 0.021943282335996628, + 0.007153339684009552, + 0.026678884401917458, + 0.01192424912005663, + 0.010864515788853168, + 0.040448930114507675, + 0.017594903707504272, + 0.024838309735059738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0016337393317371607 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.0031947072129696608 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031947072129696608, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031947072129696608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09384743869304657, + 0.10098672658205032, + 0.04208727926015854, + 0.03234945237636566, + 0.06693822890520096, + 0.02561184950172901, + 0.03217550367116928, + 0.02446778304874897, + 0.22808034718036652, + 0.35345539450645447 + ], + "credit_map": [ + 0.0063003432005643845, + 0.009527087211608887, + 0.013623172417283058, + 0.0199086032807827, + 0.016451699659228325, + 0.006691895425319672, + 0.00786169245839119, + 0.013482950627803802, + 0.007695760577917099, + 0.007417045999318361, + 0.018693696707487106, + 0.013876275159418583, + 0.020306311547756195, + 0.009261738508939743, + 0.00834664236754179, + 0.015389258041977882, + 0.0033856334630399942, + 0.0035041982773691416, + 0.004715013783425093, + 0.009874412789940834, + 0.007996216416358948, + 0.0053637223318219185, + 0.004532582592219114, + 0.002715500770136714, + 0.0025993569288402796, + 0.0019187169382348657, + 0.003868071362376213, + 0.004696562886238098, + 0.004915026947855949, + 0.005307498387992382, + 0.003195488592609763, + 0.005848732311278582, + 0.006732990499585867, + 0.0037522963248193264, + 0.02705925516784191, + 0.004171893931925297, + 0.003018694231286645, + 0.007241879589855671, + 0.005126196891069412, + 0.009835018776357174, + 0.0028122025541961193, + 0.0017758389003574848, + 0.0025162275414913893, + 0.0037851377855986357, + 0.003967692609876394, + 0.006428256630897522, + 0.0025682600680738688, + 0.0017582335276529193, + 0.0064874980598688126, + 0.0019225055584684014, + 0.010915357619524002, + 0.0021966511849313974, + 0.002348627196624875, + 0.00337370578199625, + 0.0028819998260587454, + 0.0020491578616201878, + 0.003221957478672266, + 0.001657120417803526, + 0.003643403295427561, + 0.0025225281715393066, + 0.001988859148696065, + 0.004153445363044739, + 0.0021818417590111494, + 0.00509862694889307, + 0.02977921813726425, + 0.014474373310804367, + 0.04677551984786987, + 0.01927565224468708, + 0.015688063576817513, + 0.0394841730594635, + 0.023016031831502914, + 0.039587315171957016, + 0.047577034682035446, + 0.016037913039326668, + 0.06260639429092407, + 0.026272719725966454, + 0.0239811223000288, + 0.0842025950551033, + 0.037672869861125946, + 0.05510472506284714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.0031947072129696608 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.001515312702395022 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.001515312702395022, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001515312702395022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03520898148417473, + 0.022085091099143028, + 0.015759866684675217, + 0.013821613974869251, + 0.05558827146887779, + 0.06392298638820648, + 0.08531150221824646, + 0.030664097517728806, + 0.2110169529914856, + 0.46662068367004395 + ], + "credit_map": [ + 0.0032558292150497437, + 0.0032816273160278797, + 0.005951088387519121, + 0.005092415492981672, + 0.00608939491212368, + 0.0037634256295859814, + 0.0031525574158877134, + 0.0046226452104747295, + 0.0018328613368794322, + 0.0018041931325569749, + 0.0024457487743347883, + 0.00451396219432354, + 0.004204709082841873, + 0.0025065296795219183, + 0.00241622026078403, + 0.002360865706577897, + 0.0017126675229519606, + 0.0013683284632861614, + 0.0023807224351912737, + 0.0025407609064131975, + 0.002605578163638711, + 0.002094428986310959, + 0.00177393585909158, + 0.0012834434164687991, + 0.0017513784114271402, + 0.001030385261401534, + 0.0016724636079743505, + 0.0031495217699557543, + 0.001791978138498962, + 0.0017733195563778281, + 0.0017631243681535125, + 0.000889443326741457, + 0.006594179663807154, + 0.0030094173271209, + 0.01624862290918827, + 0.003902275115251541, + 0.005552902817726135, + 0.006423286162316799, + 0.0037049150560051203, + 0.010152671486139297, + 0.007543238811194897, + 0.005387666169553995, + 0.005555205512791872, + 0.010852595791220665, + 0.011075365357100964, + 0.011029716581106186, + 0.0072058578953146935, + 0.005273337941616774, + 0.016147486865520477, + 0.004914462566375732, + 0.031969550997018814, + 0.005951027851551771, + 0.003840387100353837, + 0.008264261297881603, + 0.008025838062167168, + 0.006198483519256115, + 0.0055463556200265884, + 0.0032358034513890743, + 0.003936795052140951, + 0.0035354981664568186, + 0.002797595923766494, + 0.004020667169243097, + 0.002801616210490465, + 0.004789766389876604, + 0.029063044115900993, + 0.014967269264161587, + 0.04231660068035126, + 0.017941074445843697, + 0.01524361316114664, + 0.03440149500966072, + 0.021229468286037445, + 0.03585439920425415, + 0.063200443983078, + 0.02116720750927925, + 0.08297452330589294, + 0.03492998331785202, + 0.03185749053955078, + 0.11072412878274918, + 0.04980876296758652, + 0.07195815443992615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.001515312702395022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.043047528713941574 + ], + "opposite_logprob_tokens": [ + -3.1680474281311035 + ], + "logprob_reward": -0.043047528713941574, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043047528713941574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022521140053868294, + 0.014707068912684917, + 0.012900946661829948, + 0.02481953613460064, + 0.14638149738311768, + 0.055841464549303055, + 0.024352241307497025, + 0.009973573498427868, + 0.3014097511768341, + 0.3870927393436432 + ], + "credit_map": [ + 0.002711702138185501, + 0.0027156248688697815, + 0.0024002855643630028, + 0.0038226835895329714, + 0.004143459722399712, + 0.0015192111022770405, + 0.002475061686709523, + 0.0027331123128533363, + 0.0014420985244214535, + 0.0016845668433234096, + 0.0027113454416394234, + 0.002456164686009288, + 0.00201033940538764, + 0.0013538128696382046, + 0.001482032355852425, + 0.0015667096013203263, + 0.0015257553895935416, + 0.0012737519573420286, + 0.0015497079584747553, + 0.002647493267431855, + 0.001816432923078537, + 0.0013370781671255827, + 0.0014943048590794206, + 0.0012564216740429401, + 0.0033111388329416513, + 0.002316384809091687, + 0.0034287457820028067, + 0.0034502041526138783, + 0.0033296893816441298, + 0.004242249298840761, + 0.0021707231644541025, + 0.0025704011786729097, + 0.015920570120215416, + 0.009663686156272888, + 0.05699300393462181, + 0.008556111715734005, + 0.007752930745482445, + 0.016893045976758003, + 0.01464301347732544, + 0.01595914550125599, + 0.008317620493471622, + 0.004503509029746056, + 0.008331499062478542, + 0.008251158520579338, + 0.007499449420720339, + 0.008834021165966988, + 0.0056362454779446125, + 0.004467963241040707, + 0.0029569894541054964, + 0.004483642987906933, + 0.0037153600715100765, + 0.0035089461598545313, + 0.0029428070411086082, + 0.0030818379018455744, + 0.002028438728302717, + 0.0016342182643711567, + 0.001416134531609714, + 0.0008523511933162808, + 0.0011494540376588702, + 0.000956013158429414, + 0.0009911052184179425, + 0.0016189527232199907, + 0.000890772498678416, + 0.002098790369927883, + 0.038481902331113815, + 0.018948571756482124, + 0.059662654995918274, + 0.02426399104297161, + 0.02084321714937687, + 0.056263357400894165, + 0.0297556035220623, + 0.0531904362142086, + 0.0522165521979332, + 0.017228880897164345, + 0.06975270807743073, + 0.028956426307559013, + 0.02592887543141842, + 0.09039953351020813, + 0.0412047915160656, + 0.06140501797199249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.043047528713941574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.011074270121753216 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074270121753216, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074270121753216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13449865579605103, + 0.02892894297838211, + 0.034839946776628494, + 0.026892900466918945, + 0.12699352204799652, + 0.027843598276376724, + 0.02893191948533058, + 0.044029600918293, + 0.14526034891605377, + 0.4017805755138397 + ], + "credit_map": [ + 0.01266674418002367, + 0.010925371199846268, + 0.035759132355451584, + 0.014456147328019142, + 0.011119520291686058, + 0.007730424869805574, + 0.00793466530740261, + 0.03390665724873543, + 0.0027099191211163998, + 0.0028380437288433313, + 0.003729347139596939, + 0.006492698099464178, + 0.004079329315572977, + 0.0032196484971791506, + 0.0032030188012868166, + 0.0026569399051368237, + 0.004859883803874254, + 0.0031707666348665953, + 0.0037458522710949183, + 0.006334150210022926, + 0.005319653078913689, + 0.002957272808998823, + 0.003912659361958504, + 0.004539708141237497, + 0.002507552970200777, + 0.0016848640516400337, + 0.0035843898076564074, + 0.004992853384464979, + 0.0037348021287471056, + 0.00439666211605072, + 0.0022050293628126383, + 0.003786749206483364, + 0.009401453658938408, + 0.00746579747647047, + 0.05293460562825203, + 0.006929952651262283, + 0.004685885272920132, + 0.013090262189507484, + 0.006341181695461273, + 0.02614438720047474, + 0.003964082803577185, + 0.0017220106674358249, + 0.008372983895242214, + 0.0037638109643012285, + 0.0019474845612421632, + 0.0038622093852609396, + 0.0021521823946386576, + 0.0020588343031704426, + 0.0047906069085001945, + 0.0018641487695276737, + 0.010761917568743229, + 0.0018820263212546706, + 0.0017913178307935596, + 0.0026788481045514345, + 0.002642404055222869, + 0.0025206501595675945, + 0.005116536747664213, + 0.0038976820651441813, + 0.006227029021829367, + 0.003381879534572363, + 0.0042799776419997215, + 0.005288130138069391, + 0.004711871966719627, + 0.011126497760415077, + 0.01842081919312477, + 0.009351585060358047, + 0.02852487750351429, + 0.011607922613620758, + 0.009893094189465046, + 0.027888327836990356, + 0.013501390814781189, + 0.026072334498167038, + 0.05563288927078247, + 0.01961963251233101, + 0.07146451622247696, + 0.03137553110718727, + 0.02778874710202217, + 0.09148790687322617, + 0.044400956481695175, + 0.0600103996694088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.011074270121753216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, slowly and steadily advancing, focuses on a pale forearm of a figure, set against a soft, out-of-focus background. Soft, diffused morning light from an unseen source illuminates the scene, capturing the tiny, almost invisible hairs on the skin. A delicate network of shallow, healing wounds dot the forearm, their pale pink lines contrasting sharply with the surrounding skin. Miraculously, from these wounds, vibrant green sprouts begin to carefully break through the surface, not with force but with gentle, organic persistence. The camera's slow advance allows us to witness these incredibly fragile sprouts steadily elongate and unfold tiny, perfectly shaped leaves that shimmer as if kissed by dewdrops. The overall atmosphere is one of quiet, surreal wonder, blending the visceral reality of scarred skin with the quiet, magical beauty of new life growing through the passage of time, in a continuous and graceful movement.", + "seed": 4864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_9/index_3867_seed_4864_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": -0.026179911382496358, + "vlm_logprob": -0.026179911382496358, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.026179911382496358, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.013407780788838863 + ], + "opposite_logprob_tokens": [ + -4.6384077072143555 + ], + "logprob_reward": -0.013407780788838863, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013407780788838863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1781567484140396, + 0.03664333000779152, + 0.03129659965634346, + 0.04354654997587204, + 0.09422262758016586, + 0.07210680097341537, + 0.4049187898635864, + 0.031325243413448334, + 0.08318823575973511, + 0.024595098569989204 + ], + "credit_map": [ + 0.015070941299200058, + 0.016280557960271835, + 0.03195430710911751, + 0.018521564081311226, + 0.02152249962091446, + 0.02466391958296299, + 0.020489608868956566, + 0.029653336852788925, + 0.00377746787853539, + 0.0035292855463922024, + 0.005974334664642811, + 0.003762049600481987, + 0.004384009167551994, + 0.004495799541473389, + 0.004971609916538, + 0.005748774856328964, + 0.003088625380769372, + 0.0025190955493599176, + 0.004779641050845385, + 0.0033334768377244473, + 0.0036999990697950125, + 0.005549153313040733, + 0.003932064399123192, + 0.004394542425870895, + 0.0037776269018650055, + 0.004706557374447584, + 0.010186676867306232, + 0.00789143331348896, + 0.002565062139183283, + 0.00556959630921483, + 0.002820893656462431, + 0.006028704810887575, + 0.00785695482045412, + 0.009256605058908463, + 0.021709179505705833, + 0.0131728146225214, + 0.008845196105539799, + 0.011975377798080444, + 0.00963066890835762, + 0.011775827966630459, + 0.004864477552473545, + 0.007094226311892271, + 0.009000822901725769, + 0.006811110768467188, + 0.008821776136755943, + 0.014238771982491016, + 0.010520090349018574, + 0.010755528695881367, + 0.03332676738500595, + 0.09637326002120972, + 0.05890653654932976, + 0.07056304812431335, + 0.05140599235892296, + 0.027344131842255592, + 0.026596995070576668, + 0.04040206968784332, + 0.003510993905365467, + 0.0047066425904631615, + 0.0038189890328794718, + 0.0033520732540637255, + 0.004072775598615408, + 0.004456125665456057, + 0.00412407424300909, + 0.0032835707534104586, + 0.009202275425195694, + 0.00933137908577919, + 0.011425395496189594, + 0.008139434270560741, + 0.009484121575951576, + 0.005969575606286526, + 0.014538821764290333, + 0.015097228810191154, + 0.0031983989756554365, + 0.0033502434380352497, + 0.003933759871870279, + 0.0026895864866673946, + 0.0022169051226228476, + 0.0033980240114033222, + 0.0035469518043100834, + 0.002261227695271373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.013407780788838863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.04307744279503822 + ], + "opposite_logprob_tokens": [ + -3.1680774688720703 + ], + "logprob_reward": -0.04307744279503822, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307744279503822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17701110243797302, + 0.04949551448225975, + 0.028372449800372124, + 0.03329198807477951, + 0.10846824944019318, + 0.07897403091192245, + 0.3862285017967224, + 0.0433543436229229, + 0.07186535745859146, + 0.02293843775987625 + ], + "credit_map": [ + 0.014080340974032879, + 0.01630314812064171, + 0.031159883365035057, + 0.01759060099720955, + 0.022086411714553833, + 0.02524709515273571, + 0.020208587870001793, + 0.030335044488310814, + 0.004912443924695253, + 0.004387570079416037, + 0.01030176691710949, + 0.0054683745838701725, + 0.004756408743560314, + 0.00620005838572979, + 0.005937076639384031, + 0.0075318170711398125, + 0.002680840203538537, + 0.002289394848048687, + 0.004784435033798218, + 0.0029488448053598404, + 0.003117038868367672, + 0.00512416148558259, + 0.0033705998212099075, + 0.004057134035974741, + 0.0026426203548908234, + 0.0031026769429445267, + 0.007747111842036247, + 0.005626318510621786, + 0.0024765334092080593, + 0.004513039253652096, + 0.002657009521499276, + 0.0045266784727573395, + 0.009714462794363499, + 0.01041271723806858, + 0.02445387840270996, + 0.01482311636209488, + 0.011225126683712006, + 0.014019903726875782, + 0.010883654467761517, + 0.012935390695929527, + 0.006262109149247408, + 0.008737316355109215, + 0.007437427062541246, + 0.00749107263982296, + 0.008369343355298042, + 0.01599433831870556, + 0.013160987757146358, + 0.011521438136696815, + 0.03200539946556091, + 0.09212765097618103, + 0.05611223354935646, + 0.06719115376472473, + 0.049618326127529144, + 0.02537931129336357, + 0.025281865149736404, + 0.03851258009672165, + 0.004643536172807217, + 0.005402104929089546, + 0.007387108169496059, + 0.005782131105661392, + 0.004905261564999819, + 0.004105334635823965, + 0.005467670038342476, + 0.005661201197654009, + 0.0076815290376544, + 0.012229531072080135, + 0.009121401235461235, + 0.007644014433026314, + 0.007774593308568001, + 0.005545569118112326, + 0.011245272122323513, + 0.010623441077768803, + 0.0031970846466720104, + 0.003216338576748967, + 0.004337516613304615, + 0.002014969475567341, + 0.0019055361626669765, + 0.002757194684818387, + 0.0037674931809306145, + 0.0017423046519979835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.04307744279503822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.014203292317688465 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203292317688465, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203292317688465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1229945495724678, + 0.07231330871582031, + 0.01638777367770672, + 0.020104026421904564, + 0.09938474744558334, + 0.060441818088293076, + 0.46698081493377686, + 0.07538562268018723, + 0.04358072206377983, + 0.022426599636673927 + ], + "credit_map": [ + 0.010578534565865993, + 0.011036944575607777, + 0.026141319423913956, + 0.015361439436674118, + 0.01683536358177662, + 0.013013551011681557, + 0.015122141689062119, + 0.014905259944498539, + 0.0061765871942043304, + 0.006188904866576195, + 0.01595156453549862, + 0.009226645343005657, + 0.006056154612451792, + 0.009368585422635078, + 0.0066636293195188046, + 0.012681234627962112, + 0.002344477688893676, + 0.0015768478624522686, + 0.0020337440073490143, + 0.0017134324880316854, + 0.0018937512068077922, + 0.002487617079168558, + 0.0019458479946479201, + 0.002392055466771126, + 0.002105286344885826, + 0.0028379270806908607, + 0.002597044687718153, + 0.0038751133251935244, + 0.0021545274648815393, + 0.0018819161923602223, + 0.002850632881745696, + 0.00180157832801342, + 0.009304612874984741, + 0.009717616252601147, + 0.021119166165590286, + 0.013035669922828674, + 0.011983202770352364, + 0.011476750485599041, + 0.011946404352784157, + 0.010801319032907486, + 0.004159271251410246, + 0.005915122572332621, + 0.010118070058524609, + 0.006539073772728443, + 0.005887260194867849, + 0.009865131229162216, + 0.008386834524571896, + 0.009571054019033909, + 0.03907483071088791, + 0.11261764913797379, + 0.0680391788482666, + 0.08080020546913147, + 0.06164194643497467, + 0.028396861627697945, + 0.029870182275772095, + 0.04653992876410484, + 0.0074273403733968735, + 0.007488772738724947, + 0.013634724542498589, + 0.009450984187424183, + 0.009171664714813232, + 0.00785790104418993, + 0.010977771133184433, + 0.0093764653429389, + 0.004349524155259132, + 0.008995642885565758, + 0.005388813093304634, + 0.0056992387399077415, + 0.005177372135221958, + 0.003077106084674597, + 0.007100988179445267, + 0.0037920395843684673, + 0.0026380608323961496, + 0.003820426994934678, + 0.003520979778841138, + 0.0022072955034673214, + 0.00208586105145514, + 0.002561658388003707, + 0.0033785849809646606, + 0.0022137307096272707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.014203292317688465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.054956890642642975 + ], + "opposite_logprob_tokens": [ + -2.9299569129943848 + ], + "logprob_reward": -0.054956890642642975, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054956890642642975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0936596468091011, + 0.025907525792717934, + 0.026270106434822083, + 0.03900397568941116, + 0.07534065842628479, + 0.07993412762880325, + 0.5210351943969727, + 0.061543941497802734, + 0.05842505022883415, + 0.01887975074350834 + ], + "credit_map": [ + 0.008166376501321793, + 0.008792905136942863, + 0.016535088419914246, + 0.011048677377402782, + 0.014288319274783134, + 0.011739431880414486, + 0.010802761651575565, + 0.012286088429391384, + 0.002878030063584447, + 0.0026016603223979473, + 0.003150259843096137, + 0.002734225708991289, + 0.0029278118163347244, + 0.0036906672175973654, + 0.0034452350810170174, + 0.004479635041207075, + 0.0032450391445308924, + 0.0025124852545559406, + 0.0026224092580378056, + 0.002522022696211934, + 0.003810650436207652, + 0.004125773441046476, + 0.003002430545166135, + 0.004429296590387821, + 0.004343817476183176, + 0.0036845095455646515, + 0.008965916000306606, + 0.0054333508014678955, + 0.003886573947966099, + 0.0032695007976144552, + 0.005880923010408878, + 0.0035393827129155397, + 0.007461730856448412, + 0.00787026435136795, + 0.008255108259618282, + 0.009265334345400333, + 0.011936982162296772, + 0.009077231399714947, + 0.011960824951529503, + 0.00951317884027958, + 0.005307657178491354, + 0.007513603661209345, + 0.01785433106124401, + 0.010892868973314762, + 0.01134051289409399, + 0.006484827492386103, + 0.007332272361963987, + 0.013208052143454552, + 0.04423551261425018, + 0.1256091594696045, + 0.07530485093593597, + 0.09036308526992798, + 0.06569436937570572, + 0.03288158401846886, + 0.035586245357990265, + 0.051360394805669785, + 0.00430447468534112, + 0.010065262205898762, + 0.005849501583725214, + 0.006744750775396824, + 0.012654261663556099, + 0.006296954583376646, + 0.009320315904915333, + 0.006308419164270163, + 0.006094646640121937, + 0.008970385417342186, + 0.006206214893609285, + 0.007604519370943308, + 0.009112860076129436, + 0.0040967888198792934, + 0.010664652101695538, + 0.005674983840435743, + 0.0028611517045646906, + 0.0018176628509536386, + 0.0037073164712637663, + 0.0016238533426076174, + 0.0018977069994434714, + 0.0018142227781936526, + 0.0028899151366204023, + 0.002267920644953847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.054956890642642975 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.005254150368273258 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254150368273258, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254150368273258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0776137039065361, + 0.04776082560420036, + 0.07397781312465668, + 0.10869140177965164, + 0.0800626277923584, + 0.06293481588363647, + 0.478900671005249, + 0.03485986962914467, + 0.020571140572428703, + 0.014627112075686455 + ], + "credit_map": [ + 0.011584562249481678, + 0.011240932159125805, + 0.007992333732545376, + 0.008806675672531128, + 0.008291820995509624, + 0.010437997058033943, + 0.008603071793913841, + 0.010656309314072132, + 0.004730443935841322, + 0.004633839707821608, + 0.005736945196986198, + 0.006359664257615805, + 0.0048568639904260635, + 0.00719279982149601, + 0.004898569080978632, + 0.009351696819067001, + 0.007295446004718542, + 0.007788757793605328, + 0.008578833192586899, + 0.006804830860346556, + 0.009030556306242943, + 0.013837493024766445, + 0.007083804812282324, + 0.013558088801801205, + 0.012569984421133995, + 0.010963978245854378, + 0.012033081613481045, + 0.01212027482688427, + 0.011767450720071793, + 0.016793958842754364, + 0.01838666759431362, + 0.014056006446480751, + 0.008473314344882965, + 0.0067733838222920895, + 0.00940839946269989, + 0.009344320744276047, + 0.012500896118581295, + 0.012580162845551968, + 0.012191198766231537, + 0.008790946565568447, + 0.007565599866211414, + 0.0061091212555766106, + 0.011486021801829338, + 0.0074979630298912525, + 0.007421534042805433, + 0.006217939779162407, + 0.009094852954149246, + 0.007541784550994635, + 0.03947330266237259, + 0.11509416252374649, + 0.06961026042699814, + 0.08332154154777527, + 0.061689864844083786, + 0.03061779960989952, + 0.03155611455440521, + 0.04753764718770981, + 0.003727218834683299, + 0.003853666828945279, + 0.00559947919100523, + 0.005495133809745312, + 0.0039883386343717575, + 0.003857392817735672, + 0.004081476479768753, + 0.0042571621015667915, + 0.001914111664518714, + 0.0026514651253819466, + 0.0022708955220878124, + 0.003031262895092368, + 0.003158960025757551, + 0.0016506608808413148, + 0.003594447625800967, + 0.002299334853887558, + 0.0021856895182281733, + 0.0017747814999893308, + 0.002475681947544217, + 0.0015886785695329309, + 0.0015084079932421446, + 0.0015296241035684943, + 0.002141062170267105, + 0.0014231866225600243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.005254150368273258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, slowly and steadily advancing, focuses on a pale forearm of a figure, set against a soft, out-of-focus background. Soft, diffused morning light from an unseen source illuminates the scene, capturing the tiny, almost invisible hairs on the skin. A delicate network of shallow, healing wounds dot the forearm, their pale pink lines contrasting sharply with the surrounding skin. Miraculously, from these wounds, vibrant green sprouts begin to carefully break through the surface, not with force but with gentle, organic persistence. The camera's slow advance allows us to witness these incredibly fragile sprouts steadily elongate and unfold tiny, perfectly shaped leaves that shimmer as if kissed by dewdrops. The overall atmosphere is one of quiet, surreal wonder, blending the visceral reality of scarred skin with the quiet, magical beauty of new life growing through the passage of time, in a continuous and graceful movement.", + "seed": 4865, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000131/group_9/index_3867_seed_4865_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": -0.01475827544927597, + "vlm_logprob": -0.01475827544927597, + "vlm_yes_no_margin": 5.049999952316284, + "vlm_token_credit": -0.01475827544927597, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.002427014522254467 + ], + "opposite_logprob_tokens": [ + -7.252427101135254 + ], + "logprob_reward": -0.002427014522254467, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.002427014522254467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14617891609668732, + 0.16648134589195251, + 0.08476977795362473, + 0.09931591898202896, + 0.041466373950242996, + 0.03071439452469349, + 0.1247873604297638, + 0.04670950397849083, + 0.2250175178050995, + 0.03455881401896477 + ], + "credit_map": [ + 0.009045659564435482, + 0.010108855552971363, + 0.0529521219432354, + 0.0265902541577816, + 0.008793425746262074, + 0.014485692605376244, + 0.016191551461815834, + 0.008011356927454472, + 0.012516336515545845, + 0.006611707154661417, + 0.0709814503788948, + 0.022756554186344147, + 0.0068086665123701096, + 0.018325285986065865, + 0.020312661305069923, + 0.008168705739080906, + 0.005430134013295174, + 0.008187957108020782, + 0.021392153576016426, + 0.009980475530028343, + 0.00570807047188282, + 0.021641818806529045, + 0.006138896103948355, + 0.0062902700155973434, + 0.009600669145584106, + 0.006838405970484018, + 0.013026319444179535, + 0.009391095489263535, + 0.00557742640376091, + 0.03835991397500038, + 0.007706994656473398, + 0.008815105073153973, + 0.0027823769487440586, + 0.0023018058855086565, + 0.010424968786537647, + 0.004989412613213062, + 0.0033884034492075443, + 0.012318417429924011, + 0.002609147457405925, + 0.002651846269145608, + 0.0018575884168967605, + 0.002346493536606431, + 0.005441720597445965, + 0.004709723871201277, + 0.003017663024365902, + 0.007592412177473307, + 0.0020973689388483763, + 0.0036514245439320803, + 0.0033241952769458294, + 0.002563512185588479, + 0.04752060025930405, + 0.006451235618442297, + 0.00382597791031003, + 0.04980640113353729, + 0.004727649502456188, + 0.006567796692252159, + 0.0019650745671242476, + 0.0023609695490449667, + 0.012607447803020477, + 0.0036972048692405224, + 0.0032585286535322666, + 0.015744438394904137, + 0.0027529136277735233, + 0.0043229288421571255, + 0.006030426826328039, + 0.006133977323770523, + 0.0920877680182457, + 0.013358279131352901, + 0.008112751878798008, + 0.0690450593829155, + 0.008522418327629566, + 0.021726857870817184, + 0.002048398368060589, + 0.003447309136390686, + 0.006189631298184395, + 0.005523855332285166, + 0.0035867278929799795, + 0.004631162621080875, + 0.002588956616818905, + 0.006542775314301252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.002427014522254467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.00763415452092886 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.00763415452092886, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763415452092886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1338777393102646, + 0.3351250886917114, + 0.07992760092020035, + 0.03827308118343353, + 0.019479338079690933, + 0.014803217723965645, + 0.05784022808074951, + 0.04078992083668709, + 0.252047061920166, + 0.027836671099066734 + ], + "credit_map": [ + 0.0069349477998912334, + 0.00917768944054842, + 0.046032365411520004, + 0.02253214828670025, + 0.0077681648544967175, + 0.01960780657827854, + 0.011324241757392883, + 0.01050037145614624, + 0.02423202246427536, + 0.016414718702435493, + 0.13744068145751953, + 0.04717613756656647, + 0.013677958399057388, + 0.03850502893328667, + 0.04099487513303757, + 0.016683675348758698, + 0.004922857508063316, + 0.005209718365222216, + 0.019646313041448593, + 0.006356615107506514, + 0.00391461793333292, + 0.030242156237363815, + 0.0048669022507965565, + 0.004768421873450279, + 0.0028247905429452658, + 0.002302554901689291, + 0.005967689678072929, + 0.004008911084383726, + 0.0027272491715848446, + 0.01398329809308052, + 0.0027566382195800543, + 0.0037019518204033375, + 0.0025647280272096395, + 0.001474824850447476, + 0.0027263190131634474, + 0.0033923890441656113, + 0.002492805477231741, + 0.003041790099814534, + 0.00184758345130831, + 0.0019389003282412887, + 0.0010995841585099697, + 0.0010753496317192912, + 0.0028391892556101084, + 0.0017578037222847342, + 0.001112638390623033, + 0.004649224691092968, + 0.0009304618579335511, + 0.001338966772891581, + 0.0017118590185418725, + 0.0010880600893869996, + 0.02241058275103569, + 0.0030780902598053217, + 0.0019252178026363254, + 0.022990485653281212, + 0.0020982809364795685, + 0.0025376530829817057, + 0.001584109733812511, + 0.0016039065085351467, + 0.013539966195821762, + 0.001748488750308752, + 0.0019020996987819672, + 0.015915360301733017, + 0.001640075002796948, + 0.002855917438864708, + 0.006916080601513386, + 0.005020934622734785, + 0.10239699482917786, + 0.013356656767427921, + 0.008038386702537537, + 0.09676805883646011, + 0.009279822930693626, + 0.010270156897604465, + 0.0019978040363639593, + 0.002837102860212326, + 0.005975021980702877, + 0.004379238933324814, + 0.002202256117016077, + 0.0051076640374958515, + 0.002067490480840206, + 0.003270095447078347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.00763415452092886 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.0024909917265176773 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024909917265176773, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024909917265176773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05390120670199394, + 0.0826917216181755, + 0.04143421724438667, + 0.025852037593722343, + 0.05958489701151848, + 0.028307883068919182, + 0.14308537542819977, + 0.019539054483175278, + 0.5123967528343201, + 0.03320683166384697 + ], + "credit_map": [ + 0.003054179484024644, + 0.0065724849700927734, + 0.007613967172801495, + 0.007711910177022219, + 0.007594047114253044, + 0.010045237839221954, + 0.006554277613759041, + 0.004755096975713968, + 0.005885600578039885, + 0.004883599001914263, + 0.0339784249663353, + 0.011426343582570553, + 0.004015823360532522, + 0.008612181060016155, + 0.009230373427271843, + 0.004659367259591818, + 0.003109776182100177, + 0.004070536233484745, + 0.013742565177381039, + 0.006113420240581036, + 0.002456586342304945, + 0.005214516539126635, + 0.0038099715020507574, + 0.0029168431647121906, + 0.002862070919945836, + 0.003294948721304536, + 0.004199195187538862, + 0.00417847977951169, + 0.0028026297222822905, + 0.0041938768699765205, + 0.002162870252504945, + 0.0021579668391495943, + 0.003371627302840352, + 0.0029100272804498672, + 0.014577164314687252, + 0.005406864918768406, + 0.0037198858335614204, + 0.02130706049501896, + 0.003497365163639188, + 0.004794901702553034, + 0.0012713194591924548, + 0.0023752173874527216, + 0.005739319138228893, + 0.0030760460067540407, + 0.0023215918336063623, + 0.008634727448225021, + 0.0019213680643588305, + 0.0029682940803468227, + 0.003765277797356248, + 0.0022410228848457336, + 0.059494730085134506, + 0.007013486232608557, + 0.004097985569387674, + 0.05630267783999443, + 0.004569052252918482, + 0.005601144395768642, + 0.0017404929967597127, + 0.0010336264967918396, + 0.004256951157003641, + 0.0021314078476279974, + 0.0017622353043407202, + 0.0050328634679317474, + 0.001471957191824913, + 0.00210952153429389, + 0.013988979160785675, + 0.0075829182751476765, + 0.21425682306289673, + 0.02634337730705738, + 0.013629784807562828, + 0.19512277841567993, + 0.01826612278819084, + 0.023205969482660294, + 0.0023039565421640873, + 0.0021772426553070545, + 0.0072928243316709995, + 0.005179710686206818, + 0.0023209843784570694, + 0.005191971082240343, + 0.0027558812871575356, + 0.00598426116630435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.0024909917265176773 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.043054837733507156 + ], + "opposite_logprob_tokens": [ + -3.1680548191070557 + ], + "logprob_reward": -0.043054837733507156, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043054837733507156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056483522057533264, + 0.04321209713816643, + 0.022202806547284126, + 0.0198623426258564, + 0.03223758935928345, + 0.012041173875331879, + 0.2843649089336395, + 0.009534372016787529, + 0.48793911933898926, + 0.03212209790945053 + ], + "credit_map": [ + 0.0034890819806605577, + 0.003178816521540284, + 0.02263552136719227, + 0.008931818418204784, + 0.0022729188203811646, + 0.0057809073477983475, + 0.0068175578489899635, + 0.0033768927678465843, + 0.0028909952379763126, + 0.0032778182066977024, + 0.016747064888477325, + 0.00619655754417181, + 0.0023118881508708, + 0.00485402112826705, + 0.004552746191620827, + 0.0023810036946088076, + 0.0016807698411867023, + 0.0024245367385447025, + 0.005395348649471998, + 0.0026855245232582092, + 0.0022327792830765247, + 0.004454931244254112, + 0.0017739232862368226, + 0.0015549894887953997, + 0.001654062420129776, + 0.0021595973521471024, + 0.003308885730803013, + 0.002822571201249957, + 0.0018331031315028667, + 0.0037117532920092344, + 0.0018706831615418196, + 0.002501684706658125, + 0.0014757326571270823, + 0.0017986433813348413, + 0.007409571669995785, + 0.0038610361516475677, + 0.00287578790448606, + 0.009903390891849995, + 0.002186249243095517, + 0.0027271758299320936, + 0.0007297295960597694, + 0.0011626128107309341, + 0.002411517081782222, + 0.0021506184712052345, + 0.0017403910169377923, + 0.0017487338045611978, + 0.0010753085371106863, + 0.001022261567413807, + 0.0068763243034482, + 0.003932182211428881, + 0.11590379476547241, + 0.012841328047215939, + 0.007572907023131847, + 0.11750917881727219, + 0.009128398261964321, + 0.010600773617625237, + 0.0009062052704393864, + 0.0007964054821059108, + 0.0020805385429412127, + 0.001193291274830699, + 0.0014337751781567931, + 0.0015994436107575893, + 0.0007343000615946949, + 0.000790411839261651, + 0.011927645653486252, + 0.007054077927023172, + 0.2052842378616333, + 0.024265479296445847, + 0.011544734239578247, + 0.19097867608070374, + 0.01701025292277336, + 0.019873974844813347, + 0.001907972851768136, + 0.0022096107713878155, + 0.008476674556732178, + 0.00321710086427629, + 0.001867553568445146, + 0.005239630118012428, + 0.0023626601323485374, + 0.006840887945145369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.043054837733507156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 focusing on a pale human forearm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shallow, healing wounds visible on the forearm as pale pink lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do vibrant green sprouts break through the skin from the wounds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily advance towards the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background soft and out-of-focus with diffused morning 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": [ + -0.018184378743171692 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.018184378743171692, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018184378743171692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2453906387090683, + 0.10863235592842102, + 0.05616150051355362, + 0.025434838607907295, + 0.0231099221855402, + 0.021522710099816322, + 0.10827614367008209, + 0.10759242624044418, + 0.2564547061920166, + 0.047424785792827606 + ], + "credit_map": [ + 0.012723040767014027, + 0.028766531497240067, + 0.03461002558469772, + 0.0432346872985363, + 0.022063879296183586, + 0.05995370075106621, + 0.019791511818766594, + 0.024247245863080025, + 0.007313621696084738, + 0.008968722075223923, + 0.03140052780508995, + 0.019141850993037224, + 0.007567683234810829, + 0.01562613993883133, + 0.010741041041910648, + 0.007872764021158218, + 0.0024131713435053825, + 0.0027649139519780874, + 0.018647965043783188, + 0.003865545615553856, + 0.0026461235247552395, + 0.019514452666044235, + 0.00333379115909338, + 0.0029755381401628256, + 0.0014132432406768203, + 0.001948332181200385, + 0.005896014627069235, + 0.004259311594069004, + 0.002334750257432461, + 0.005680169444531202, + 0.0020821515936404467, + 0.0018208653200417757, + 0.001900346833281219, + 0.0015295416815206409, + 0.0049604494124650955, + 0.0030490169301629066, + 0.0023228249046951532, + 0.0056244442239403725, + 0.001743606640957296, + 0.0019796923734247684, + 0.0018878401024267077, + 0.0016658519161865115, + 0.004966075997799635, + 0.003114065621048212, + 0.001981594366952777, + 0.004883510060608387, + 0.001401911024004221, + 0.0016218593809753656, + 0.006640290841460228, + 0.0073404451832175255, + 0.03536108136177063, + 0.008836448192596436, + 0.009189395233988762, + 0.023887725546956062, + 0.004989537876099348, + 0.012031219899654388, + 0.004355862736701965, + 0.00462655583396554, + 0.03477197512984276, + 0.006671423092484474, + 0.005008190404623747, + 0.039236802607774734, + 0.005377193912863731, + 0.00754442298784852, + 0.006725134328007698, + 0.004748285282403231, + 0.10271302610635757, + 0.012872900813817978, + 0.007661266252398491, + 0.10296056419610977, + 0.009234691970050335, + 0.00953883770853281, + 0.0042206402868032455, + 0.0029129674658179283, + 0.008403059095144272, + 0.006451194640249014, + 0.005073601845651865, + 0.011072681285440922, + 0.004300897009670734, + 0.004989744629710913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.018184378743171692 + } + ], + "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/0000146/group_1/index_tensor([4905, 4905])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_tensor([4905, 4905])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b31f7dd9da3c4dca39f7e59a79cf0284f86da9d6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_tensor([4905, 4905])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, extreme close-up focuses intently on a pair of hands leisurely manipulating a tape measure on a dark, grainy wooden workbench. A single, warm light source from above casts a focused pool of illumination, creating sharp highlights on joints and metal components while casting the surrounding area into soft shadow. The hands belong to a man, clean, yet revealing the faint callouses of manual labor. One hand firmly grasps the black and silver plastic casing of the tape measure, which bears slight wear and scratches from frequent use. The thumb of the other hand hooks the metal end tab of the tape and carefully pulls outward, gently curving the vibrant yellow, slightly curved metal ribbon. The tape extends steadily for several inches, its bold black measurement markings becoming visible before the fingers pause, then suddenly release. With a rapid mechanical whir, an internal spring retracts the ribbon, which disappears back into the casing until the metal tab strikes the plastic with a sharp, unmistakable click, creating a tense, rhythmic tension before the hands resume their idle motion.", + "seed": 467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_4905_seed_467_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": -0.022533435747027396, + "vlm_logprob": -0.022533435747027396, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.022533435747027396, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02094823122024536 + ], + "opposite_logprob_tokens": [ + -5.02094841003418 + ], + "logprob_reward": -0.02094823122024536, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.02094823122024536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306343674659729, + 0.24505169689655304, + 0.09406153112649918, + 0.0914049968123436, + 0.037617601454257965, + 0.05520060285925865, + 0.06565728783607483, + 0.04871015250682831, + 0.06779299676418304, + 0.1638687551021576 + ], + "credit_map": [ + 0.015195033513009548, + 0.010693821124732494, + 0.024067608639597893, + 0.014085507020354271, + 0.01781218871474266, + 0.01126245129853487, + 0.020797496661543846, + 0.016720255836844444, + 0.06016968935728073, + 0.015034987591207027, + 0.026878345757722855, + 0.014354053884744644, + 0.01645011641085148, + 0.015276570804417133, + 0.015686439350247383, + 0.08120150119066238, + 0.009822359308600426, + 0.0070065720938146114, + 0.014788974076509476, + 0.00987411942332983, + 0.01266599539667368, + 0.007960853166878223, + 0.01657659187912941, + 0.015366069041192532, + 0.011244588531553745, + 0.010431621223688126, + 0.015320277772843838, + 0.008542614057660103, + 0.013902957551181316, + 0.011789832264184952, + 0.00902811624109745, + 0.011144997552037239, + 0.004523615352809429, + 0.003667138982564211, + 0.004287144169211388, + 0.003665083786472678, + 0.007874484173953533, + 0.003643456846475601, + 0.0050527057610452175, + 0.0049039744772017, + 0.008184491656720638, + 0.004526018630713224, + 0.008013867773115635, + 0.006487557198852301, + 0.00412004766985774, + 0.008292579092085361, + 0.007077129557728767, + 0.008498910814523697, + 0.006125668529421091, + 0.005232698284089565, + 0.008702059276401997, + 0.006260262802243233, + 0.008472442626953125, + 0.01116697583347559, + 0.010032899677753448, + 0.009664284996688366, + 0.006067320704460144, + 0.0035917486529797316, + 0.007174039725214243, + 0.0043287742882966995, + 0.0042149098590016365, + 0.004930194932967424, + 0.008079901337623596, + 0.010323269292712212, + 0.008969031274318695, + 0.0083234878256917, + 0.007695064414292574, + 0.007861546240746975, + 0.009981216862797737, + 0.008405492641031742, + 0.009372023865580559, + 0.007185137830674648, + 0.017828384414315224, + 0.01582682505249977, + 0.0129829291254282, + 0.01309139747172594, + 0.024045508354902267, + 0.030967839062213898, + 0.02293749339878559, + 0.02618839405477047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.02094823122024536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014224916696548462 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.014224916696548462, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014224916696548462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10455320030450821, + 0.43563538789749146, + 0.054311491549015045, + 0.04370195418596268, + 0.03214562311768532, + 0.027229072526097298, + 0.06446565687656403, + 0.06011594459414482, + 0.10958877950906754, + 0.06825288385152817 + ], + "credit_map": [ + 0.007987207733094692, + 0.006907079368829727, + 0.012304369360208511, + 0.015168514102697372, + 0.014662273228168488, + 0.006601064000278711, + 0.02661835588514805, + 0.014304328709840775, + 0.10326140373945236, + 0.023370498791337013, + 0.04685661196708679, + 0.028972622007131577, + 0.03431597352027893, + 0.030793821439146996, + 0.02718815766274929, + 0.14087632298469543, + 0.008301724679768085, + 0.0033486555330455303, + 0.009462915360927582, + 0.007176120765507221, + 0.008945602923631668, + 0.003623370546847582, + 0.007848730310797691, + 0.005604371894150972, + 0.008643927052617073, + 0.005330298095941544, + 0.008540187031030655, + 0.0031284757424145937, + 0.005218443926423788, + 0.003314156783744693, + 0.004548815079033375, + 0.0049776481464505196, + 0.0028946734964847565, + 0.003217375371605158, + 0.0038422425277531147, + 0.004286449868232012, + 0.00678990688174963, + 0.0033874313812702894, + 0.003954155836254358, + 0.003773386124521494, + 0.005009776912629604, + 0.0022348465863615274, + 0.0038578740786761045, + 0.0028421797323971987, + 0.002702372847124934, + 0.0032519877422600985, + 0.002970615401864052, + 0.004359417594969273, + 0.00411062641069293, + 0.00297693838365376, + 0.01017975527793169, + 0.0038178504910320044, + 0.007617796305567026, + 0.013509003445506096, + 0.0084104398265481, + 0.013843243941664696, + 0.005579231306910515, + 0.005614118184894323, + 0.008857457898557186, + 0.007359924726188183, + 0.005272453650832176, + 0.005967096891254187, + 0.01079507451504469, + 0.010670590214431286, + 0.01331078540533781, + 0.010063973255455494, + 0.01187026035040617, + 0.013198217377066612, + 0.015322865918278694, + 0.016898954287171364, + 0.016365477815270424, + 0.012558244168758392, + 0.0075706481002271175, + 0.006565055809915066, + 0.006200892850756645, + 0.006379144731909037, + 0.009257573634386063, + 0.014098511077463627, + 0.009887946769595146, + 0.008293111808598042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.014224916696548462 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097023263573647 + ], + "opposite_logprob_tokens": [ + -4.511096954345703 + ], + "logprob_reward": -0.011097023263573647, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097023263573647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1082073226571083, + 0.0778617113828659, + 0.06186230480670929, + 0.09700040519237518, + 0.06808584928512573, + 0.2446909248828888, + 0.14706557989120483, + 0.037918977439403534, + 0.05821269378066063, + 0.09909422695636749 + ], + "credit_map": [ + 0.01388439629226923, + 0.007916160859167576, + 0.021987590938806534, + 0.013188072480261326, + 0.012520904652774334, + 0.008426370099186897, + 0.016735026612877846, + 0.013548786751925945, + 0.01676057279109955, + 0.0048250616528093815, + 0.009343953803181648, + 0.00485214265063405, + 0.006797480396926403, + 0.005386537406593561, + 0.007963183335959911, + 0.021932771429419518, + 0.0066272420808672905, + 0.004269821103662252, + 0.012663361616432667, + 0.0061096493154764175, + 0.007130345795303583, + 0.006031130440533161, + 0.009650793857872486, + 0.009379959665238857, + 0.014323834329843521, + 0.01012030802667141, + 0.016692131757736206, + 0.008837624453008175, + 0.009374691173434258, + 0.013699527829885483, + 0.010700645856559277, + 0.013251638039946556, + 0.010327561758458614, + 0.006340827327221632, + 0.011946803890168667, + 0.006490695755928755, + 0.007950222119688988, + 0.007334006018936634, + 0.008678053505718708, + 0.00901767611503601, + 0.05367673188447952, + 0.02502519264817238, + 0.03159131482243538, + 0.02159142680466175, + 0.01944269798696041, + 0.02586359903216362, + 0.026265762746334076, + 0.04123419523239136, + 0.009816623292863369, + 0.005071950610727072, + 0.03190615773200989, + 0.007654045708477497, + 0.008562728762626648, + 0.036187317222356796, + 0.011473052203655243, + 0.03639369457960129, + 0.0040162247605621815, + 0.0029482587706297636, + 0.0058332523331046104, + 0.0041867815889418125, + 0.004110151901841164, + 0.0038455950561910868, + 0.006938067730516195, + 0.0060406457632780075, + 0.00920968409627676, + 0.0069406963884830475, + 0.005254484713077545, + 0.004505952354520559, + 0.0101638063788414, + 0.00872438307851553, + 0.0064711179584264755, + 0.0069425636902451515, + 0.009049080312252045, + 0.009749913588166237, + 0.008186929859220982, + 0.010769089683890343, + 0.014053719118237495, + 0.017980575561523438, + 0.013531890697777271, + 0.015773024410009384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.011097023263573647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023298438638448715 + ], + "opposite_logprob_tokens": [ + -3.773298501968384 + ], + "logprob_reward": -0.023298438638448715, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023298438638448715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06828851252794266, + 0.09282739460468292, + 0.027588268741965294, + 0.22137314081192017, + 0.048691339790821075, + 0.06991355121135712, + 0.08243471384048462, + 0.045954857021570206, + 0.18854154646396637, + 0.15438665449619293 + ], + "credit_map": [ + 0.006329915951937437, + 0.004883411340415478, + 0.007241858635097742, + 0.007528702728450298, + 0.012475074268877506, + 0.0036376784555613995, + 0.015505307354032993, + 0.010686565190553665, + 0.021879486739635468, + 0.005371136590838432, + 0.010122370906174183, + 0.00566093809902668, + 0.007137408945709467, + 0.007554170209914446, + 0.005932455416768789, + 0.02916943095624447, + 0.0034599562641233206, + 0.002060175407677889, + 0.0038210575003176928, + 0.0031063882634043694, + 0.0033664656803011894, + 0.0024100542068481445, + 0.004701710771769285, + 0.004662461578845978, + 0.02754150703549385, + 0.026580827310681343, + 0.03523118793964386, + 0.020327985286712646, + 0.03428447246551514, + 0.028466807678341866, + 0.021316317841410637, + 0.02762402407824993, + 0.005908194929361343, + 0.0045754569582641125, + 0.007409827318042517, + 0.00391206843778491, + 0.00790714006870985, + 0.005604347214102745, + 0.006312394980341196, + 0.007061914075165987, + 0.012888966128230095, + 0.006285729352384806, + 0.010025270283222198, + 0.005373817402869463, + 0.006080800201743841, + 0.009477450512349606, + 0.0060082366690039635, + 0.013773281127214432, + 0.011601477861404419, + 0.007549505680799484, + 0.011464263312518597, + 0.008411831222474575, + 0.010487113147974014, + 0.011318965815007687, + 0.012473070062696934, + 0.009128491394221783, + 0.005822800565510988, + 0.004741339944303036, + 0.004871779587119818, + 0.004766268655657768, + 0.003517098957672715, + 0.004286137875169516, + 0.007111302576959133, + 0.010838126763701439, + 0.021033722907304764, + 0.022679409012198448, + 0.022953899577260017, + 0.017984852194786072, + 0.030317697674036026, + 0.03182723745703697, + 0.021118057891726494, + 0.02062666043639183, + 0.013928504660725594, + 0.01169767789542675, + 0.01094788033515215, + 0.014131243340671062, + 0.019990017637610435, + 0.029562989249825478, + 0.02159157581627369, + 0.03253678232431412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.023298438638448715 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430985689163208 + ], + "opposite_logprob_tokens": [ + -3.1680984497070312 + ], + "logprob_reward": -0.0430985689163208, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430985689163208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07336092740297318, + 0.4059896767139435, + 0.035937950015068054, + 0.03027515672147274, + 0.06956488639116287, + 0.08140373229980469, + 0.08987312018871307, + 0.02671108953654766, + 0.06269457191228867, + 0.12418880313634872 + ], + "credit_map": [ + 0.0055620805360376835, + 0.006185418926179409, + 0.013285870663821697, + 0.008567108772695065, + 0.010684976354241371, + 0.0058064148761332035, + 0.013551358133554459, + 0.00971770565956831, + 0.10893192142248154, + 0.021418355405330658, + 0.04573157802224159, + 0.017503587529063225, + 0.02311769872903824, + 0.02235320582985878, + 0.024950427934527397, + 0.14198294281959534, + 0.005745451431721449, + 0.0029703115578740835, + 0.006788987200707197, + 0.0035949074663221836, + 0.003965006209909916, + 0.0028146756812930107, + 0.005169700365513563, + 0.00488891638815403, + 0.003812227863818407, + 0.003679092740640044, + 0.003818703116849065, + 0.003242124803364277, + 0.003881708951666951, + 0.004593439865857363, + 0.003252329071983695, + 0.003995534498244524, + 0.0062477425672113895, + 0.005448103882372379, + 0.008973908610641956, + 0.009413995780050755, + 0.015048294328153133, + 0.0073347389698028564, + 0.009615125134587288, + 0.007482981309294701, + 0.010234948247671127, + 0.007857718504965305, + 0.011523307301104069, + 0.009350983425974846, + 0.010094190016388893, + 0.011195152997970581, + 0.008383766748011112, + 0.012763673439621925, + 0.0056234076619148254, + 0.003356156637892127, + 0.01868751086294651, + 0.005399279296398163, + 0.004935015924274921, + 0.020562827587127686, + 0.00813321303576231, + 0.023175718262791634, + 0.004025923553854227, + 0.0025132375303655863, + 0.002567540854215622, + 0.0025893566198647022, + 0.0021659445483237505, + 0.0022414112463593483, + 0.004805582109838724, + 0.005802095402032137, + 0.0065275635570287704, + 0.005643602926284075, + 0.0059624831192195415, + 0.005906723439693451, + 0.014032548293471336, + 0.008741315454244614, + 0.008955704048275948, + 0.006924642249941826, + 0.014081100933253765, + 0.012186207808554173, + 0.012039939872920513, + 0.016132842749357224, + 0.016940992325544357, + 0.02478891611099243, + 0.013115407899022102, + 0.014903406612575054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.0430985689163208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme close-up focuses intently on a pair of hands leisurely manipulating a tape measure on a dark, grainy wooden workbench. A single, warm light source from above casts a focused pool of illumination, creating sharp highlights on joints and metal components while casting the surrounding area into soft shadow. The hands belong to a man, clean, yet revealing the faint callouses of manual labor. One hand firmly grasps the black and silver plastic casing of the tape measure, which bears slight wear and scratches from frequent use. The thumb of the other hand hooks the metal end tab of the tape and carefully pulls outward, gently curving the vibrant yellow, slightly curved metal ribbon. The tape extends steadily for several inches, its bold black measurement markings becoming visible before the fingers pause, then suddenly release. With a rapid mechanical whir, an internal spring retracts the ribbon, which disappears back into the casing until the metal tab strikes the plastic with a sharp, unmistakable click, creating a tense, rhythmic tension before the hands resume their idle motion.", + "seed": 468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_4905_seed_468_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": -0.023443502094596624, + "vlm_logprob": -0.023443502094596624, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.023443502094596624, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018432049080729485 + ], + "opposite_logprob_tokens": [ + -4.643432140350342 + ], + "logprob_reward": -0.018432049080729485, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.018432049080729485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1969270259141922, + 0.05563447251915932, + 0.03613992780447006, + 0.027239782735705376, + 0.020381178706884384, + 0.02262507937848568, + 0.03673822805285454, + 0.11540193110704422, + 0.27515703439712524, + 0.21375541388988495 + ], + "credit_map": [ + 0.0330672524869442, + 0.023965217173099518, + 0.04304058104753494, + 0.017999913543462753, + 0.016175784170627594, + 0.017290379852056503, + 0.018286829814314842, + 0.027101056650280952, + 0.008938808925449848, + 0.011025134474039078, + 0.00821306649595499, + 0.00668980972841382, + 0.0049881055019795895, + 0.005231925286352634, + 0.004593460354954004, + 0.005954159423708916, + 0.006335769779980183, + 0.008549273945391178, + 0.004740524105727673, + 0.0038333330303430557, + 0.0026818460319191217, + 0.002878208877518773, + 0.003033276414498687, + 0.004087693989276886, + 0.004613934550434351, + 0.0024112598039209843, + 0.004603346809744835, + 0.003071762388572097, + 0.002993060275912285, + 0.0025050051044672728, + 0.0037104615475982428, + 0.0033309522550553083, + 0.003460140433162451, + 0.0037941141054034233, + 0.00314535410143435, + 0.002307202899828553, + 0.0016475172014907002, + 0.0017317271558567882, + 0.002328472211956978, + 0.0019666512962430716, + 0.0034113104920834303, + 0.002171600703150034, + 0.003970024175941944, + 0.0022893769200891256, + 0.002473284024745226, + 0.0023065009154379368, + 0.002849712036550045, + 0.0031532696448266506, + 0.006151361856609583, + 0.006249831058084965, + 0.004933521151542664, + 0.0035469369031488895, + 0.004048162139952183, + 0.0029407849069684744, + 0.004076894372701645, + 0.004790736362338066, + 0.013817198574543, + 0.0121882613748312, + 0.01421622559428215, + 0.021981418132781982, + 0.014180488884449005, + 0.012089666910469532, + 0.011548185721039772, + 0.015380487777292728, + 0.02769455313682556, + 0.0287022702395916, + 0.03966357931494713, + 0.044091030955314636, + 0.033497076481580734, + 0.035173479467630386, + 0.03598102182149887, + 0.030354032292962074, + 0.015674317255616188, + 0.03557679057121277, + 0.037524715065956116, + 0.021600116044282913, + 0.02645588293671608, + 0.025188850238919258, + 0.019385457038879395, + 0.032349277287721634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.018432049080729485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609804667532444 + ], + "opposite_logprob_tokens": [ + -4.1410980224609375 + ], + "logprob_reward": -0.01609804667532444, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609804667532444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10981419682502747, + 0.11839805543422699, + 0.043312039226293564, + 0.014842947013676167, + 0.016317218542099, + 0.01650220714509487, + 0.026899399235844612, + 0.06706628203392029, + 0.26823297142982483, + 0.3186146914958954 + ], + "credit_map": [ + 0.018974026665091515, + 0.010418727062642574, + 0.020190883427858353, + 0.012073157355189323, + 0.009752349928021431, + 0.01240967120975256, + 0.01230580359697342, + 0.013689580373466015, + 0.019729673862457275, + 0.024903863668441772, + 0.017935683950781822, + 0.01350824348628521, + 0.00966812577098608, + 0.00831687729805708, + 0.010964336805045605, + 0.0133712412789464, + 0.00729031627997756, + 0.008975948207080364, + 0.006701681297272444, + 0.005124679300934076, + 0.0035905621480196714, + 0.002735947724431753, + 0.003939643502235413, + 0.004953260999172926, + 0.002469404600560665, + 0.0022380670998245478, + 0.002799237845465541, + 0.0016569782746955752, + 0.0010626172879710793, + 0.0011848160065710545, + 0.0012267724378034472, + 0.0022050526458770037, + 0.002663109917193651, + 0.0019957374315708876, + 0.002753588603809476, + 0.0022009103558957577, + 0.0013650035252794623, + 0.0012461020378395915, + 0.0017273430712521076, + 0.0023654247634112835, + 0.00276655750349164, + 0.002166236052289605, + 0.003005390288308263, + 0.0018503830069676042, + 0.0016536335460841656, + 0.0014351505087688565, + 0.0016764646861702204, + 0.0019483903888612986, + 0.004557077307254076, + 0.0032223055604845285, + 0.004610867239534855, + 0.004387679044157267, + 0.0022662756964564323, + 0.0022129060234874487, + 0.002305807778611779, + 0.0033364822156727314, + 0.008889054879546165, + 0.008139761164784431, + 0.011496243998408318, + 0.01155568566173315, + 0.006213918793946505, + 0.004944241605699062, + 0.0065792277455329895, + 0.009248143993318081, + 0.031452372670173645, + 0.04054651036858559, + 0.04261001572012901, + 0.040809422731399536, + 0.026611031964421272, + 0.022530989721417427, + 0.02279086224734783, + 0.04088177904486656, + 0.031123731285333633, + 0.03391411527991295, + 0.10003712028265, + 0.028915446251630783, + 0.02697164937853813, + 0.019961915910243988, + 0.020626479759812355, + 0.057064227759838104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.01609804667532444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640188094228506 + ], + "opposite_logprob_tokens": [ + -4.882640361785889 + ], + "logprob_reward": -0.007640188094228506, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640188094228506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08410823345184326, + 0.066035695374012, + 0.03943055123090744, + 0.020287752151489258, + 0.028945719823241234, + 0.04231248423457146, + 0.04833420366048813, + 0.1867634803056717, + 0.19418515264987946, + 0.2895967364311218 + ], + "credit_map": [ + 0.012771952897310257, + 0.008738495409488678, + 0.014333062805235386, + 0.011984066106379032, + 0.00849884282797575, + 0.0068285660818219185, + 0.011135745793581009, + 0.009817499667406082, + 0.009782725013792515, + 0.008315867744386196, + 0.010932429693639278, + 0.007975063286721706, + 0.007477108854800463, + 0.0063484711572527885, + 0.0079525550827384, + 0.0072514768689870834, + 0.006027928553521633, + 0.004147854167968035, + 0.006659698206931353, + 0.004532970488071442, + 0.004251326899975538, + 0.0036967655178159475, + 0.005295837298035622, + 0.004818168003112078, + 0.002983751939609647, + 0.0024021025747060776, + 0.0027280962094664574, + 0.002517497166991234, + 0.0021520487498492002, + 0.0021182887721806765, + 0.002800973830744624, + 0.002584991278126836, + 0.0038376539014279842, + 0.003254431765526533, + 0.004885637667030096, + 0.003991683479398489, + 0.0032965848222374916, + 0.0029445940162986517, + 0.0033704019151628017, + 0.003364734584465623, + 0.006656294222921133, + 0.005517494399100542, + 0.007299505174160004, + 0.004036689642816782, + 0.0042299809865653515, + 0.003784661181271076, + 0.005433863028883934, + 0.005353996064513922, + 0.0074637760408222675, + 0.004429744556546211, + 0.008362570777535439, + 0.005570972803980112, + 0.004941257182508707, + 0.004775389563292265, + 0.007416110951453447, + 0.005374385043978691, + 0.030034223571419716, + 0.018636206164956093, + 0.02979959547519684, + 0.024734344333410263, + 0.022428108379244804, + 0.01700153946876526, + 0.020636025816202164, + 0.023493440821766853, + 0.019395358860492706, + 0.024168269708752632, + 0.026706796139478683, + 0.030113792046904564, + 0.020526546984910965, + 0.01865050196647644, + 0.026706675067543983, + 0.027917208150029182, + 0.028348801657557487, + 0.026385219767689705, + 0.08736786991357803, + 0.02526191994547844, + 0.021925847977399826, + 0.03062783181667328, + 0.016676481813192368, + 0.053002770990133286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.007640188094228506 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635827101767063 + ], + "opposite_logprob_tokens": [ + -3.6513583660125732 + ], + "logprob_reward": -0.02635827101767063, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635827101767063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09834353625774384, + 0.07347210496664047, + 0.03334563598036766, + 0.026995189487934113, + 0.03218560665845871, + 0.01426247414201498, + 0.03568458929657936, + 0.15537479519844055, + 0.14103087782859802, + 0.3893052041530609 + ], + "credit_map": [ + 0.015089981257915497, + 0.013269064947962761, + 0.015181014314293861, + 0.014017429202795029, + 0.006911251228302717, + 0.006907885894179344, + 0.008035574108362198, + 0.018931331112980843, + 0.011788953095674515, + 0.013743117451667786, + 0.011137201450765133, + 0.009054219350218773, + 0.006596013437956572, + 0.005770613439381123, + 0.006896161008626223, + 0.008485819213092327, + 0.005469996947795153, + 0.00626082718372345, + 0.005184408742934465, + 0.0032779457978904247, + 0.003310864558443427, + 0.002550530480220914, + 0.0034391821827739477, + 0.003851876128464937, + 0.00428277300670743, + 0.004034131299704313, + 0.004636789672076702, + 0.003850183915346861, + 0.0024651275016367435, + 0.0023078948725014925, + 0.0025853211991488934, + 0.0028329642955213785, + 0.005290750879794359, + 0.003419141750782728, + 0.0052434601821005344, + 0.0030880665872246027, + 0.0033918297849595547, + 0.0038108467124402523, + 0.0035605086013674736, + 0.00438099866732955, + 0.002504973439499736, + 0.0024515073746442795, + 0.002399198245257139, + 0.0017625691834837198, + 0.0012655210448428988, + 0.0011203419417142868, + 0.0013114199973642826, + 0.0014469412853941321, + 0.005181519780308008, + 0.004116362426429987, + 0.006084635388106108, + 0.0055596306920051575, + 0.0033778557553887367, + 0.0023436350747942924, + 0.004618796985596418, + 0.004402149468660355, + 0.021327786147594452, + 0.017785243690013885, + 0.026615533977746964, + 0.02955642342567444, + 0.011895381845533848, + 0.010716023854911327, + 0.017960716038942337, + 0.019517676904797554, + 0.015401029027998447, + 0.011384733021259308, + 0.04550356790423393, + 0.01437405776232481, + 0.006482036784291267, + 0.00702389981597662, + 0.012613398022949696, + 0.028248142451047897, + 0.04118160903453827, + 0.06811508536338806, + 0.07892420887947083, + 0.04121562838554382, + 0.03958738595247269, + 0.024073699489235878, + 0.044361136853694916, + 0.05184644088149071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.02635827101767063 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868895560503006 + ], + "opposite_logprob_tokens": [ + -3.0486888885498047 + ], + "logprob_reward": -0.04868895560503006, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868895560503006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30963724851608276, + 0.09212550520896912, + 0.03379522264003754, + 0.022266389802098274, + 0.025733526796102524, + 0.022837160155177116, + 0.018954332917928696, + 0.09737136214971542, + 0.15468645095825195, + 0.2225927859544754 + ], + "credit_map": [ + 0.04785705730319023, + 0.03397652134299278, + 0.04988192394375801, + 0.02061910182237625, + 0.037706390023231506, + 0.033411771059036255, + 0.03717857599258423, + 0.04900592193007469, + 0.014981841668486595, + 0.008435185067355633, + 0.013570720329880714, + 0.0064072092063724995, + 0.010994416661560535, + 0.012304402887821198, + 0.012784160673618317, + 0.012647559866309166, + 0.005280980374664068, + 0.003016421804204583, + 0.005116146057844162, + 0.0024755848571658134, + 0.004207845311611891, + 0.004108300432562828, + 0.004746031481772661, + 0.004843914415687323, + 0.003644494106993079, + 0.0022352945525199175, + 0.0031422425527125597, + 0.0017148080514743924, + 0.0027205138467252254, + 0.002702427329495549, + 0.0032954011112451553, + 0.0028112069703638554, + 0.004321810323745012, + 0.002554058562964201, + 0.0036932395305484533, + 0.0019929083064198494, + 0.0032183120492845774, + 0.003137465799227357, + 0.003997175954282284, + 0.0028185569681227207, + 0.003945451695472002, + 0.00206959736533463, + 0.0035422754008322954, + 0.001996474340558052, + 0.002703329548239708, + 0.002413123147562146, + 0.003739949781447649, + 0.0024269591085612774, + 0.003237758297473192, + 0.0026084044948220253, + 0.002788340672850609, + 0.001982763409614563, + 0.002147004008293152, + 0.0017521762056276202, + 0.002128173131495714, + 0.0023097125813364983, + 0.015010440722107887, + 0.011862529441714287, + 0.01615728624165058, + 0.011185403913259506, + 0.010703034698963165, + 0.009074627421796322, + 0.011042884550988674, + 0.012335155159235, + 0.017447730526328087, + 0.015193001367151737, + 0.03281642124056816, + 0.01979175955057144, + 0.011189598590135574, + 0.019674593582749367, + 0.015255659818649292, + 0.023317690938711166, + 0.020539885386824608, + 0.02965068444609642, + 0.062411826103925705, + 0.01693693734705448, + 0.019236668944358826, + 0.020964300259947777, + 0.012679560109972954, + 0.040172915905714035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.04868895560503006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme close-up focuses intently on a pair of hands leisurely manipulating a tape measure on a dark, grainy wooden workbench. A single, warm light source from above casts a focused pool of illumination, creating sharp highlights on joints and metal components while casting the surrounding area into soft shadow. The hands belong to a man, clean, yet revealing the faint callouses of manual labor. One hand firmly grasps the black and silver plastic casing of the tape measure, which bears slight wear and scratches from frequent use. The thumb of the other hand hooks the metal end tab of the tape and carefully pulls outward, gently curving the vibrant yellow, slightly curved metal ribbon. The tape extends steadily for several inches, its bold black measurement markings becoming visible before the fingers pause, then suddenly release. With a rapid mechanical whir, an internal spring retracts the ribbon, which disappears back into the casing until the metal tab strikes the plastic with a sharp, unmistakable click, creating a tense, rhythmic tension before the hands resume their idle motion.", + "seed": 469, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_4905_seed_469_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": -0.031557540595531466, + "vlm_logprob": -0.031557540595531466, + "vlm_yes_no_margin": 3.699999952316284, + "vlm_token_credit": -0.031557540595531466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02286871150135994 + ], + "opposite_logprob_tokens": [ + -4.772868633270264 + ], + "logprob_reward": -0.02286871150135994, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.02286871150135994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.184477761387825, + 0.05012892931699753, + 0.0675426572561264, + 0.07393056154251099, + 0.035190753638744354, + 0.034951549023389816, + 0.04659293591976166, + 0.19521741569042206, + 0.16674315929412842, + 0.14522427320480347 + ], + "credit_map": [ + 0.018592078238725662, + 0.013776462525129318, + 0.03873886540532112, + 0.027015672996640205, + 0.014534637331962585, + 0.023789267987012863, + 0.017373286187648773, + 0.030657490715384483, + 0.006705178413540125, + 0.006521327421069145, + 0.008283646777272224, + 0.00668426975607872, + 0.003243451938033104, + 0.005093109328299761, + 0.005075276363641024, + 0.008522666990756989, + 0.009293950162827969, + 0.006130095571279526, + 0.01511269249022007, + 0.009855263866484165, + 0.003947337158024311, + 0.004353356547653675, + 0.008166201412677765, + 0.010683758184313774, + 0.008343877270817757, + 0.006660277955234051, + 0.01301205437630415, + 0.009761068038642406, + 0.004780855495482683, + 0.007998393848538399, + 0.011427291668951511, + 0.01194674801081419, + 0.00499305734410882, + 0.0027313181199133396, + 0.005312787368893623, + 0.0063161239959299564, + 0.002676690462976694, + 0.004973620641976595, + 0.0026458685752004385, + 0.005541286896914244, + 0.004233772400766611, + 0.003845792729407549, + 0.0061362977139651775, + 0.004609484691172838, + 0.0026254283729940653, + 0.005280201323330402, + 0.003228374058380723, + 0.004992198199033737, + 0.007190439850091934, + 0.00463769119232893, + 0.006713977549225092, + 0.004685716237872839, + 0.004586585331708193, + 0.006926648784428835, + 0.005449201911687851, + 0.006402673665434122, + 0.029321709647774696, + 0.021676545962691307, + 0.029198428615927696, + 0.021276947110891342, + 0.014221337623894215, + 0.026684407144784927, + 0.02871551550924778, + 0.02412252314388752, + 0.04074662923812866, + 0.020267387852072716, + 0.01571163535118103, + 0.018517229706048965, + 0.017264636233448982, + 0.017217865213751793, + 0.015206429176032543, + 0.021811343729496002, + 0.017907869070768356, + 0.02714141458272934, + 0.01839888095855713, + 0.01885192096233368, + 0.019056035205721855, + 0.012984005734324455, + 0.012011782266199589, + 0.01887235976755619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02286871150135994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029841480776667595 + ], + "opposite_logprob_tokens": [ + -3.529841423034668 + ], + "logprob_reward": -0.029841480776667595, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029841480776667595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15783564746379852, + 0.15183618664741516, + 0.11920897662639618, + 0.09323220700025558, + 0.07204524427652359, + 0.0498586967587471, + 0.029530156403779984, + 0.0591750368475914, + 0.0820939689874649, + 0.18518385291099548 + ], + "credit_map": [ + 0.022716112434864044, + 0.017955949530005455, + 0.02401251159608364, + 0.019194014370441437, + 0.00912713073194027, + 0.017444176599383354, + 0.019348038360476494, + 0.028037723153829575, + 0.018722645938396454, + 0.020165756344795227, + 0.018385492265224457, + 0.022527216002345085, + 0.012222793884575367, + 0.01749284379184246, + 0.01933680661022663, + 0.022982630878686905, + 0.012173681519925594, + 0.014114857651293278, + 0.018739832565188408, + 0.018422311171889305, + 0.00857221707701683, + 0.012311081402003765, + 0.015153967775404453, + 0.019721034914255142, + 0.00968035589903593, + 0.010946984402835369, + 0.01681651920080185, + 0.014722377993166447, + 0.005791091360151768, + 0.008984343148767948, + 0.010821121744811535, + 0.015469415113329887, + 0.010174009948968887, + 0.007022821344435215, + 0.014312321320176125, + 0.011381378397345543, + 0.004358096979558468, + 0.0057171061635017395, + 0.007463551126420498, + 0.011615961790084839, + 0.007420255336910486, + 0.004857297986745834, + 0.008569547906517982, + 0.007332764100283384, + 0.003491768380627036, + 0.004602581262588501, + 0.0056781708262860775, + 0.007906311191618443, + 0.00470107002183795, + 0.0031922408379614353, + 0.004235133994370699, + 0.003255405928939581, + 0.002089268760755658, + 0.0032758936285972595, + 0.004666141699999571, + 0.004115001298487186, + 0.006411892827600241, + 0.008119460195302963, + 0.00960524007678032, + 0.006236018612980843, + 0.006839214824140072, + 0.008503512479364872, + 0.0068666222505271435, + 0.0065930746495723724, + 0.018283989280462265, + 0.011962485499680042, + 0.008831629529595375, + 0.008058127015829086, + 0.0071614524349570274, + 0.009118913672864437, + 0.0075347949750721455, + 0.011142577044665813, + 0.023614177480340004, + 0.03313149884343147, + 0.01383900549262762, + 0.017137786373496056, + 0.03409143537282944, + 0.024178506806492805, + 0.016236165538430214, + 0.0229552760720253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.029841480776667595 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821867562830448 + ], + "opposite_logprob_tokens": [ + -4.018218517303467 + ], + "logprob_reward": -0.01821867562830448, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821867562830448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15336382389068604, + 0.0877012237906456, + 0.08699996024370193, + 0.02853272296488285, + 0.04828496649861336, + 0.05724921450018883, + 0.034523505717515945, + 0.1013730838894844, + 0.14558592438697815, + 0.25638556480407715 + ], + "credit_map": [ + 0.02249477617442608, + 0.01414317823946476, + 0.022420991212129593, + 0.012714112177491188, + 0.013191699050366879, + 0.021073155105113983, + 0.022745033726096153, + 0.02458086423575878, + 0.013077845796942711, + 0.007643352262675762, + 0.015296904370188713, + 0.01085872482508421, + 0.006846404634416103, + 0.008145466446876526, + 0.01119374018162489, + 0.014638782478868961, + 0.00917002372443676, + 0.006368658505380154, + 0.022745396941900253, + 0.012955792248249054, + 0.005317900329828262, + 0.005956724751740694, + 0.009039606899023056, + 0.01544585544615984, + 0.0035670301876962185, + 0.002769548213109374, + 0.005117069464176893, + 0.004698744509369135, + 0.0018337987130507827, + 0.0027311730664223433, + 0.0031758632976561785, + 0.004639497492462397, + 0.005623888224363327, + 0.0036265829112380743, + 0.007083327509462833, + 0.009325869381427765, + 0.0026929385494440794, + 0.003074375446885824, + 0.008741645142436028, + 0.008116341196000576, + 0.006912637501955032, + 0.0056435102596879005, + 0.009949577040970325, + 0.008598352782428265, + 0.004530440550297499, + 0.006162965204566717, + 0.006191168446093798, + 0.00926055945456028, + 0.0038570913020521402, + 0.004014253616333008, + 0.004393218085169792, + 0.0041706496849656105, + 0.0036364977713674307, + 0.004536561202257872, + 0.005413905251771212, + 0.00450132554396987, + 0.012497982941567898, + 0.012998382560908794, + 0.012308403849601746, + 0.010911693796515465, + 0.009797146543860435, + 0.017798272892832756, + 0.014057813212275505, + 0.011003393679857254, + 0.02417730912566185, + 0.02247403748333454, + 0.015479376539587975, + 0.019213052466511726, + 0.015257997438311577, + 0.016536414623260498, + 0.013763315044343472, + 0.018684428185224533, + 0.035216230899095535, + 0.04181564226746559, + 0.01973925530910492, + 0.011934703215956688, + 0.08171813935041428, + 0.02501603402197361, + 0.012416700832545757, + 0.02852885238826275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01821867562830448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812139853835106 + ], + "opposite_logprob_tokens": [ + -3.288121461868286 + ], + "logprob_reward": -0.03812139853835106, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812139853835106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09365604817867279, + 0.04680655151605606, + 0.031591784209012985, + 0.04264594614505768, + 0.018767081201076508, + 0.031202275305986404, + 0.0505291149020195, + 0.24757538735866547, + 0.21984116733074188, + 0.21738460659980774 + ], + "credit_map": [ + 0.010277677327394485, + 0.009397444315254688, + 0.01890707015991211, + 0.011296979151666164, + 0.00785780604928732, + 0.014012647792696953, + 0.008339069783687592, + 0.013567361980676651, + 0.00619314331561327, + 0.0034399915020912886, + 0.007990392856299877, + 0.006388647947460413, + 0.0029911689925938845, + 0.0046777185052633286, + 0.007137817330658436, + 0.007987677119672298, + 0.002862856723368168, + 0.0016003026394173503, + 0.006077157333493233, + 0.0050981310196220875, + 0.0023945511784404516, + 0.002454162109643221, + 0.005856717471033335, + 0.005247908644378185, + 0.004767496604472399, + 0.002822075504809618, + 0.008800916373729706, + 0.005694203544408083, + 0.003260059515014291, + 0.004355512093752623, + 0.006289832293987274, + 0.006655854173004627, + 0.0020973910577595234, + 0.0019365776097401977, + 0.0022931641433387995, + 0.002407555468380451, + 0.0015998155577108264, + 0.002575790975242853, + 0.0029235915280878544, + 0.002933196723461151, + 0.0032435073517262936, + 0.00375646841712296, + 0.0042651062831282616, + 0.005263994447886944, + 0.002374103292822838, + 0.004750322084873915, + 0.0030062629375606775, + 0.004542514216154814, + 0.006277897860854864, + 0.0041635166853666306, + 0.008458275347948074, + 0.009640367701649666, + 0.0037101844791322947, + 0.005915933754295111, + 0.006147176958620548, + 0.006215768400579691, + 0.0320124514400959, + 0.03039308823645115, + 0.024551624432206154, + 0.03171520307660103, + 0.024982677772641182, + 0.043725453317165375, + 0.02999071031808853, + 0.03020419552922249, + 0.045117635279893875, + 0.029521511867642403, + 0.022916890680789948, + 0.028134280815720558, + 0.018543895334005356, + 0.026489311829209328, + 0.021220209077000618, + 0.02789745107293129, + 0.027807679027318954, + 0.03441036120057106, + 0.023171037435531616, + 0.023826634511351585, + 0.04195389151573181, + 0.018070723861455917, + 0.016574349254369736, + 0.0315699465572834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.03812139853835106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04873743653297424 + ], + "opposite_logprob_tokens": [ + -3.0487375259399414 + ], + "logprob_reward": -0.04873743653297424, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04873743653297424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2069225162267685, + 0.1359253078699112, + 0.03732948750257492, + 0.04901960492134094, + 0.06990711390972137, + 0.14042097330093384, + 0.07191598415374756, + 0.07732785493135452, + 0.09594382345676422, + 0.11528730392456055 + ], + "credit_map": [ + 0.0197162888944149, + 0.02330857329070568, + 0.02544167824089527, + 0.020083744078874588, + 0.020757809281349182, + 0.04358057677745819, + 0.022621354088187218, + 0.03141249716281891, + 0.012349903583526611, + 0.016308387741446495, + 0.0168046522885561, + 0.015868594869971275, + 0.013232271187007427, + 0.028172219172120094, + 0.013705864548683167, + 0.01948343589901924, + 0.0032024558167904615, + 0.0037415153346955776, + 0.005216141697019339, + 0.007368110120296478, + 0.002705251332372427, + 0.004214110318571329, + 0.004904840141534805, + 0.005977064371109009, + 0.003929442260414362, + 0.004431138280779123, + 0.0050183008424937725, + 0.006726135965436697, + 0.004681569989770651, + 0.007086377125233412, + 0.009420262649655342, + 0.007726376410573721, + 0.007854163646697998, + 0.007223112042993307, + 0.010353533551096916, + 0.00876200757920742, + 0.006513004191219807, + 0.01100906915962696, + 0.007305330131202936, + 0.010886900126934052, + 0.01234858576208353, + 0.014643147587776184, + 0.01931135356426239, + 0.014395017176866531, + 0.015031703747808933, + 0.02603437379002571, + 0.017788413912057877, + 0.020868370309472084, + 0.007647174410521984, + 0.009211840108036995, + 0.008157439529895782, + 0.008313125930726528, + 0.007061994168907404, + 0.010936918668448925, + 0.011873239651322365, + 0.00871424749493599, + 0.008793598972260952, + 0.009189998731017113, + 0.008081663399934769, + 0.012368149124085903, + 0.008083694614470005, + 0.009684312157332897, + 0.011064582504332066, + 0.010061854496598244, + 0.01851644553244114, + 0.013185416348278522, + 0.01259112823754549, + 0.0098463986068964, + 0.010704157873988152, + 0.012379162944853306, + 0.007109947968274355, + 0.011611169204115868, + 0.01700393296778202, + 0.015189481899142265, + 0.008633391931653023, + 0.013938592746853828, + 0.021800274029374123, + 0.012090866453945637, + 0.00855460949242115, + 0.018076149746775627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04873743653297424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, extreme close-up focuses intently on a pair of hands leisurely manipulating a tape measure on a dark, grainy wooden workbench. A single, warm light source from above casts a focused pool of illumination, creating sharp highlights on joints and metal components while casting the surrounding area into soft shadow. The hands belong to a man, clean, yet revealing the faint callouses of manual labor. One hand firmly grasps the black and silver plastic casing of the tape measure, which bears slight wear and scratches from frequent use. The thumb of the other hand hooks the metal end tab of the tape and carefully pulls outward, gently curving the vibrant yellow, slightly curved metal ribbon. The tape extends steadily for several inches, its bold black measurement markings becoming visible before the fingers pause, then suddenly release. With a rapid mechanical whir, an internal spring retracts the ribbon, which disappears back into the casing until the metal tab strikes the plastic with a sharp, unmistakable click, creating a tense, rhythmic tension before the hands resume their idle motion.", + "seed": 470, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_1/index_4905_seed_470_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": -0.02279082853347063, + "vlm_logprob": -0.02279082853347063, + "vlm_yes_no_margin": 4.1999999523162845, + "vlm_token_credit": -0.02279082853347063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016648409888148308 + ], + "opposite_logprob_tokens": [ + -5.516648292541504 + ], + "logprob_reward": -0.016648409888148308, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.016648409888148308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11454647779464722, + 0.07316507399082184, + 0.02621794305741787, + 0.1049559935927391, + 0.07034271210432053, + 0.04268093779683113, + 0.09284486621618271, + 0.08102437853813171, + 0.15580548346042633, + 0.2384161651134491 + ], + "credit_map": [ + 0.018225401639938354, + 0.011687453836202621, + 0.015726972371339798, + 0.016880527138710022, + 0.011528887785971165, + 0.010884002782404423, + 0.011616822332143784, + 0.017996391281485558, + 0.01850719004869461, + 0.0035282974131405354, + 0.012971504591405392, + 0.00577588239684701, + 0.004013746976852417, + 0.005010804161429405, + 0.005583460442721844, + 0.017774179577827454, + 0.005308516789227724, + 0.0023782476782798767, + 0.003702332731336355, + 0.004295702092349529, + 0.001980394357815385, + 0.002339476952329278, + 0.0027531443629413843, + 0.0034601257648319006, + 0.015657704323530197, + 0.008421630598604679, + 0.018055638298392296, + 0.01607506349682808, + 0.008491208776831627, + 0.004911987576633692, + 0.012478803284466267, + 0.02086395025253296, + 0.009589185006916523, + 0.006115851458162069, + 0.014064780436456203, + 0.00796769093722105, + 0.008423225954174995, + 0.005013260059058666, + 0.008870793506503105, + 0.01029791496694088, + 0.005703164264559746, + 0.004490159451961517, + 0.006368311587721109, + 0.006339536514133215, + 0.004066468682140112, + 0.003205617656931281, + 0.00630510039627552, + 0.006202576216310263, + 0.013494586572051048, + 0.005494359415024519, + 0.013277565129101276, + 0.00877884216606617, + 0.0060063800774514675, + 0.006404300685971975, + 0.0136117497459054, + 0.02577708289027214, + 0.009019631892442703, + 0.00839791726320982, + 0.009949058294296265, + 0.01575731858611107, + 0.008776303380727768, + 0.00714391702786088, + 0.008966143243014812, + 0.013014083728194237, + 0.01537266280502081, + 0.014880536124110222, + 0.019053490832448006, + 0.027747321873903275, + 0.012793623842298985, + 0.02817893587052822, + 0.022732142359018326, + 0.01504675205796957, + 0.02971871756017208, + 0.03430395573377609, + 0.05778156593441963, + 0.03625509887933731, + 0.026385297998785973, + 0.017948979511857033, + 0.019145086407661438, + 0.016877437010407448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.016648409888148308 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018207672983407974 + ], + "opposite_logprob_tokens": [ + -4.018207550048828 + ], + "logprob_reward": -0.018207672983407974, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018207672983407974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11308984458446503, + 0.10867831856012344, + 0.05803808942437172, + 0.04991768300533295, + 0.22146569192409515, + 0.18925991654396057, + 0.09215453267097473, + 0.0572548545897007, + 0.039433956146240234, + 0.07070715725421906 + ], + "credit_map": [ + 0.02141941897571087, + 0.01250297948718071, + 0.017370114102959633, + 0.016648802906274796, + 0.010290322825312614, + 0.006452804431319237, + 0.013292822986841202, + 0.015112579800188541, + 0.01781029999256134, + 0.01021628174930811, + 0.016219183802604675, + 0.013026905246078968, + 0.012354588136076927, + 0.012884709984064102, + 0.01229579746723175, + 0.013870556838810444, + 0.012001865543425083, + 0.004261791240423918, + 0.009753672406077385, + 0.007005611434578896, + 0.005630706436932087, + 0.0065474784933030605, + 0.005890753120183945, + 0.006946211215108633, + 0.008246001787483692, + 0.006188568193465471, + 0.00713730975985527, + 0.00771492812782526, + 0.005064285360276699, + 0.003688739612698555, + 0.006404678337275982, + 0.005473169963806868, + 0.02642030641436577, + 0.022974180057644844, + 0.029354432597756386, + 0.035295020788908005, + 0.033633727580308914, + 0.018664676696062088, + 0.0258799958974123, + 0.029243366792798042, + 0.030192609876394272, + 0.018833674490451813, + 0.029339658096432686, + 0.024227222427725792, + 0.01859232969582081, + 0.011222658678889275, + 0.031969036906957626, + 0.024882715195417404, + 0.012457512319087982, + 0.0077172801829874516, + 0.012821055948734283, + 0.018400415778160095, + 0.00941349845379591, + 0.006859642453491688, + 0.008798808790743351, + 0.015686318278312683, + 0.006867189891636372, + 0.007560882717370987, + 0.007516020908951759, + 0.010501276701688766, + 0.005165777634829283, + 0.006364061962813139, + 0.006373684387654066, + 0.006905958987772465, + 0.005372579209506512, + 0.00436977157369256, + 0.004914708435535431, + 0.00750828068703413, + 0.003369438461959362, + 0.0040566083043813705, + 0.005333853419870138, + 0.004508716519922018, + 0.0066185020841658115, + 0.008698055520653725, + 0.009672965854406357, + 0.01362145971506834, + 0.005917546339333057, + 0.006804549600929022, + 0.010113712400197983, + 0.009260370396077633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.018207672983407974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11878400295972824, + 0.045346084982156754, + 0.05122056230902672, + 0.04194960743188858, + 0.05535236746072769, + 0.09877806156873703, + 0.06329513341188431, + 0.046103280037641525, + 0.1168162152171135, + 0.36235469579696655 + ], + "credit_map": [ + 0.03107506036758423, + 0.008739977143704891, + 0.02119426243007183, + 0.016863485798239708, + 0.008600521832704544, + 0.008282228372991085, + 0.009442895650863647, + 0.014585559256374836, + 0.011159569025039673, + 0.0025629554875195026, + 0.008351982571184635, + 0.006912292912602425, + 0.0028618667274713516, + 0.0032555381767451763, + 0.0040544625371694565, + 0.006187415216118097, + 0.01079284306615591, + 0.00418479461222887, + 0.007204227149486542, + 0.005146529991179705, + 0.004955558571964502, + 0.004997506272047758, + 0.004741847515106201, + 0.009197249077260494, + 0.007348908577114344, + 0.00603968882933259, + 0.006192326080054045, + 0.003977181389927864, + 0.005613086279481649, + 0.004151511937379837, + 0.0036735464818775654, + 0.004953355528414249, + 0.007347496692091227, + 0.0052697728388011456, + 0.008287336677312851, + 0.008788030594587326, + 0.007088195066899061, + 0.004735954105854034, + 0.005172848701477051, + 0.00866273045539856, + 0.011591874063014984, + 0.011963846161961555, + 0.014369460754096508, + 0.018975531682372093, + 0.011218860745429993, + 0.0069893235340714455, + 0.012563971802592278, + 0.011105179786682129, + 0.010429991409182549, + 0.004135314840823412, + 0.00937223806977272, + 0.008352736011147499, + 0.004846764262765646, + 0.004306230694055557, + 0.007451667916029692, + 0.014400187879800797, + 0.0059259566478431225, + 0.006226448342204094, + 0.006464740261435509, + 0.009924592450261116, + 0.0032725606579333544, + 0.004320176318287849, + 0.004545246250927448, + 0.005423553287982941, + 0.015981098636984825, + 0.011093714274466038, + 0.01449055690318346, + 0.018217619508504868, + 0.01482093520462513, + 0.012220155447721481, + 0.01756955496966839, + 0.012422577477991581, + 0.03583535552024841, + 0.0690893828868866, + 0.06391006708145142, + 0.04141650348901749, + 0.03423720598220825, + 0.0375252328813076, + 0.03889068588614464, + 0.04145022854208946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.009802537970244884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589079707860947 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589079707860947, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589079707860947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12673823535442352, + 0.059398725628852844, + 0.04437268152832985, + 0.03042776696383953, + 0.058055102825164795, + 0.04344622418284416, + 0.07411030679941177, + 0.07997149974107742, + 0.25818315148353577, + 0.2252962738275528 + ], + "credit_map": [ + 0.02742951177060604, + 0.01233728602528572, + 0.020015813410282135, + 0.011801445856690407, + 0.009733532555401325, + 0.00843776110559702, + 0.012571928091347218, + 0.02441095933318138, + 0.009393044747412205, + 0.004546632524579763, + 0.00843223836272955, + 0.009146254509687424, + 0.00474238907918334, + 0.00476014195010066, + 0.0054282196797430515, + 0.012949803844094276, + 0.00893335696309805, + 0.0036014458164572716, + 0.006626212038099766, + 0.005380972288548946, + 0.003008951898664236, + 0.002835008082911372, + 0.00436012027785182, + 0.009626617655158043, + 0.004168218467384577, + 0.004405544605106115, + 0.004448857624083757, + 0.003629349172115326, + 0.0036641606129705906, + 0.0018031836953014135, + 0.0031938774045556784, + 0.005114574450999498, + 0.009742917492985725, + 0.005086206831037998, + 0.011390367522835732, + 0.006084538996219635, + 0.005684642121195793, + 0.004744743462651968, + 0.0057143718004226685, + 0.009607313200831413, + 0.005775137338787317, + 0.004108770750463009, + 0.006984115112572908, + 0.00577788008376956, + 0.005311408080160618, + 0.0034959737677127123, + 0.005272252950817347, + 0.006720683537423611, + 0.010687190108001232, + 0.005621317774057388, + 0.010533270426094532, + 0.013072618283331394, + 0.006303269881755114, + 0.006342748645693064, + 0.00970991887152195, + 0.011839976534247398, + 0.00947155524045229, + 0.010464360937476158, + 0.009341548196971416, + 0.011060450226068497, + 0.007722469978034496, + 0.008906529285013676, + 0.01013389602303505, + 0.012870696373283863, + 0.029979515820741653, + 0.024824155494570732, + 0.03217969089746475, + 0.06119002029299736, + 0.023033201694488525, + 0.03018282912671566, + 0.0325859934091568, + 0.02420775778591633, + 0.02863946557044983, + 0.025367848575115204, + 0.03766390681266785, + 0.044920727610588074, + 0.021591320633888245, + 0.019844813272356987, + 0.024558288976550102, + 0.022709904238581657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020589079707860947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, high-angle extreme close-up of a pair of hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the hands manipulating a black and silver tape measure on a dark, grainy wooden workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single warm light source from above create sharp highlights and soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the yellow metal tape extending, displaying black markings, and then retracting with a click?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the hands appear clean but show faint callouses of manual labor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870644211769104 + ], + "opposite_logprob_tokens": [ + -3.048706531524658 + ], + "logprob_reward": -0.04870644211769104, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870644211769104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11870864033699036, + 0.06855911761522293, + 0.06137804687023163, + 0.08360666036605835, + 0.14403197169303894, + 0.0479084849357605, + 0.10545022785663605, + 0.1114504411816597, + 0.1020062118768692, + 0.15690021216869354 + ], + "credit_map": [ + 0.021074041724205017, + 0.008735308423638344, + 0.01765548624098301, + 0.01546680647879839, + 0.009748319163918495, + 0.007727101910859346, + 0.01473647728562355, + 0.023565098643302917, + 0.015307926572859287, + 0.006662560161203146, + 0.010744530707597733, + 0.007743532303720713, + 0.006050055846571922, + 0.004474988207221031, + 0.007860677316784859, + 0.009714845567941666, + 0.017385071143507957, + 0.003610861487686634, + 0.010477564297616482, + 0.004788398742675781, + 0.003251912770792842, + 0.0026482162065804005, + 0.004854763858020306, + 0.014361263252794743, + 0.014894229359924793, + 0.007348756305873394, + 0.01373025681823492, + 0.005954799707978964, + 0.007072580512613058, + 0.004141087178140879, + 0.0074035669676959515, + 0.023061389103531837, + 0.021520309150218964, + 0.011313959024846554, + 0.029034869745373726, + 0.017712198197841644, + 0.01022148784250021, + 0.010960147716104984, + 0.014682206325232983, + 0.02858678437769413, + 0.008620765060186386, + 0.004432084038853645, + 0.00892292894423008, + 0.0060583967715501785, + 0.004268258344382048, + 0.002858572406694293, + 0.005740725435316563, + 0.007006755564361811, + 0.014439390040934086, + 0.009516716934740543, + 0.016477292403578758, + 0.01425427570939064, + 0.012960819527506828, + 0.01306748203933239, + 0.00999642163515091, + 0.01473782304674387, + 0.012192392721772194, + 0.011884360574185848, + 0.0139029361307621, + 0.01817082054913044, + 0.014046808704733849, + 0.01344110444188118, + 0.012540383264422417, + 0.015271637588739395, + 0.010909873060882092, + 0.00875143427401781, + 0.013527353294193745, + 0.013651152141392231, + 0.0108155757188797, + 0.010968269780278206, + 0.017516832798719406, + 0.01586572267115116, + 0.014282813295722008, + 0.03122680075466633, + 0.02942766807973385, + 0.02216947451233864, + 0.013940589502453804, + 0.01409794483333826, + 0.016659099608659744, + 0.015095819719135761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.04870644211769104 + } + ], + "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/0000146/group_15/index_tensor([4269, 4269])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_15/index_tensor([4269, 4269])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5aceab13c14961424c5fa9b62b2e5acf974ec95c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_15/index_tensor([4269, 4269])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a level angle, the camera held still, completely directing the viewer's attention to a vibrant moment of nature. A small sparrow perched securely on a rough, mossy apple branch. Its fluffy feathers took on a rich gradation in the light, gradating from a deep brown at the top to a pale grayish-white at the belly. Its small, obsidian-bright eyes fixed intently on the object before them: a large, plump, red apple, its skin smooth as wax, an alluring crimson with hints of yellow-green. Warm afternoon sunlight filtered through the lush green leaves overhead, casting vibrant, fragmented light spots on the bird's back, the curved surface of the apple, and the mottled branches. The sparrow's movements are imbued with rhythm. Quickly and precisely, it lowers its head, its sharp beak firmly pecking at the apple. Each peck lifts up a tiny bit of glistening flesh, revealing the juicy, glistening white flesh at the tiny opening. The entire scene is enveloped in a tranquil and focused atmosphere, dominated by warm golden sunlight, vibrant red, and vibrant green, creating a vibrant, serene, and beautiful pastoral scene.", + "seed": 6991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_15/index_4269_seed_6991_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": -0.133418532833457, + "vlm_logprob": -0.133418532833457, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.133418532833457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.258810430765152 + ], + "opposite_logprob_tokens": [ + -1.5088104009628296 + ], + "logprob_reward": -0.258810430765152, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.258810430765152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30846524238586426, + 0.12105585634708405, + 0.07173340767621994, + 0.11214645206928253, + 0.1277424842119217, + 0.11284662783145905, + 0.049788136035203934, + 0.044278088957071304, + 0.0376850925385952, + 0.014258582144975662 + ], + "credit_map": [ + 0.032596513628959656, + 0.0552092008292675, + 0.04556717723608017, + 0.032399822026491165, + 0.035726193338632584, + 0.03416107967495918, + 0.02732270024716854, + 0.04548254236578941, + 0.01118526328355074, + 0.01551034301519394, + 0.01555856503546238, + 0.01389237493276596, + 0.016743145883083344, + 0.012711780145764351, + 0.016721392050385475, + 0.018732992932200432, + 0.00885592121630907, + 0.009097949601709843, + 0.010521997697651386, + 0.008320122957229614, + 0.007228670176118612, + 0.007699840236455202, + 0.00734825199469924, + 0.012660659849643707, + 0.011336891911923885, + 0.013313326984643936, + 0.011226731352508068, + 0.0159425251185894, + 0.011838299222290516, + 0.011528220027685165, + 0.017795829102396965, + 0.01916462928056717, + 0.012607178650796413, + 0.011447074823081493, + 0.017854126170277596, + 0.01761886663734913, + 0.014581366442143917, + 0.013829992152750492, + 0.021234657615423203, + 0.01856922172009945, + 0.008691566996276379, + 0.010986235924065113, + 0.012328187935054302, + 0.01630338653922081, + 0.018529098480939865, + 0.01333251316100359, + 0.021660491824150085, + 0.011015149764716625, + 0.004561128560453653, + 0.005574802868068218, + 0.006585577502846718, + 0.006983034312725067, + 0.00457827327772975, + 0.00518376799300313, + 0.009566064924001694, + 0.006755485199391842, + 0.004055917728692293, + 0.005857319105416536, + 0.005823823157697916, + 0.00609491067007184, + 0.00523423682898283, + 0.0045230393297970295, + 0.005426791496574879, + 0.007262049708515406, + 0.004517551511526108, + 0.0030953800305724144, + 0.004756886512041092, + 0.005278684664517641, + 0.0031355309765785933, + 0.004415271338075399, + 0.004712392576038837, + 0.00777339655905962, + 0.0014117032987996936, + 0.001535377698019147, + 0.0015822341665625572, + 0.002145878504961729, + 0.0018829682376235723, + 0.001351621001958847, + 0.0024044306483119726, + 0.0019443691708147526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.258810430765152 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903086394071579 + ], + "opposite_logprob_tokens": [ + -2.579030752182007 + ], + "logprob_reward": -0.07903086394071579, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903086394071579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48586317896842957, + 0.1761951893568039, + 0.06092575937509537, + 0.05451357364654541, + 0.02844322845339775, + 0.07946255058050156, + 0.04801911115646362, + 0.023854190483689308, + 0.028660910204052925, + 0.014062311500310898 + ], + "credit_map": [ + 0.05055111274123192, + 0.0930461436510086, + 0.08945435285568237, + 0.050256822258234024, + 0.04500918835401535, + 0.053427040576934814, + 0.035459887236356735, + 0.06865864992141724, + 0.01586000993847847, + 0.02240617573261261, + 0.0257614366710186, + 0.022006336599588394, + 0.02310906909406185, + 0.01962437480688095, + 0.02915632165968418, + 0.018271462991833687, + 0.005989967379719019, + 0.006838275585323572, + 0.009102484211325645, + 0.00844001118093729, + 0.007154977880418301, + 0.005843108054250479, + 0.009374326094985008, + 0.008182613179087639, + 0.004637389909476042, + 0.007779643405228853, + 0.003915229346603155, + 0.007784446235746145, + 0.007326147984713316, + 0.005962812341749668, + 0.009862752631306648, + 0.00724515225738287, + 0.0037288765888661146, + 0.0036626129876822233, + 0.003226136788725853, + 0.004409980494529009, + 0.0028699010144919157, + 0.0029376589227467775, + 0.0036892276257276535, + 0.003918836358934641, + 0.011836547404527664, + 0.010828228667378426, + 0.015782563015818596, + 0.009301071055233479, + 0.006571442354470491, + 0.008545015938580036, + 0.007187518756836653, + 0.009410165250301361, + 0.004915389697998762, + 0.005439102184027433, + 0.008411683142185211, + 0.005841939244419336, + 0.004482558462768793, + 0.005526387598365545, + 0.005627159029245377, + 0.007774895057082176, + 0.002539036562666297, + 0.0028026066720485687, + 0.0025721644051373005, + 0.003499248530715704, + 0.0026236188132315874, + 0.002606976544484496, + 0.003822451224550605, + 0.003388086799532175, + 0.0026042365934699774, + 0.0021879086270928383, + 0.005656369496136904, + 0.004426391329616308, + 0.0031557169277220964, + 0.0027937344275414944, + 0.004581737332046032, + 0.0032548161689192057, + 0.0013694306835532188, + 0.0013376310234889388, + 0.0012166093802079558, + 0.0019210901809856296, + 0.002269257325679064, + 0.001473903888836503, + 0.0023497336078435183, + 0.002124655991792679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07903086394071579 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786011651158333 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009786011651158333, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786011651158333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2321428507566452, + 0.1844244748353958, + 0.06883727014064789, + 0.08446004986763, + 0.08109291642904282, + 0.0900420993566513, + 0.09719117730855942, + 0.08436253666877747, + 0.04978445917367935, + 0.02766212821006775 + ], + "credit_map": [ + 0.026017554104328156, + 0.03832152113318443, + 0.03408215567469597, + 0.028729230165481567, + 0.02013295143842697, + 0.024936115369200706, + 0.029330352321267128, + 0.03059297800064087, + 0.020486783236265182, + 0.029802381992340088, + 0.021623898297548294, + 0.0203521940857172, + 0.02271987870335579, + 0.02016511932015419, + 0.026861807331442833, + 0.022412413731217384, + 0.005436406470835209, + 0.006220066454261541, + 0.006476107053458691, + 0.01024157926440239, + 0.01005892176181078, + 0.007335441652685404, + 0.012739984318614006, + 0.010328765027225018, + 0.008991695940494537, + 0.013560707680881023, + 0.010014723055064678, + 0.010388629510998726, + 0.008529185317456722, + 0.009483461268246174, + 0.011092054657638073, + 0.012399588711559772, + 0.008462343364953995, + 0.0074922158382833, + 0.01042637787759304, + 0.008158300071954727, + 0.009808415547013283, + 0.00875736866146326, + 0.012245910242199898, + 0.01574198715388775, + 0.005873031914234161, + 0.007974614389240742, + 0.010624310933053493, + 0.012222294695675373, + 0.014530433341860771, + 0.01029373612254858, + 0.017786197364330292, + 0.010737476870417595, + 0.010060985572636127, + 0.011264386586844921, + 0.021771039813756943, + 0.01230037771165371, + 0.007845040410757065, + 0.010509839281439781, + 0.0108463941141963, + 0.012593111954629421, + 0.007940724492073059, + 0.006819195579737425, + 0.00803777202963829, + 0.012130792252719402, + 0.011044477112591267, + 0.009656121022999287, + 0.014821016229689121, + 0.013912437483668327, + 0.005478601437062025, + 0.006241186056286097, + 0.0063094403594732285, + 0.006060905754566193, + 0.005975889973342419, + 0.004812956787645817, + 0.008093293756246567, + 0.006812189240008593, + 0.0026710128877311945, + 0.0030147985089570284, + 0.0032383608631789684, + 0.002971869194880128, + 0.0035835695452988148, + 0.0029940716922283173, + 0.004716735798865557, + 0.004471709951758385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009786011651158333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813659906387329 + ], + "opposite_logprob_tokens": [ + -1.406365990638733 + ], + "logprob_reward": -0.2813659906387329, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813659906387329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29901114106178284, + 0.12110991030931473, + 0.08272062242031097, + 0.0846022367477417, + 0.06672990322113037, + 0.0735042616724968, + 0.11471488326787949, + 0.07460483908653259, + 0.04668937996029854, + 0.03631278872489929 + ], + "credit_map": [ + 0.03334410488605499, + 0.04337238147854805, + 0.04753066971898079, + 0.03806249052286148, + 0.029800016433000565, + 0.03154623508453369, + 0.037647511810064316, + 0.03770776465535164, + 0.014143993146717548, + 0.012727358378469944, + 0.019818812608718872, + 0.013856196776032448, + 0.014888539910316467, + 0.012662316672503948, + 0.012658167630434036, + 0.020354533568024635, + 0.010428287088871002, + 0.00817269179970026, + 0.010576254688203335, + 0.011101172305643559, + 0.00868268497288227, + 0.008104236796498299, + 0.008784791454672813, + 0.016870511695742607, + 0.00902217160910368, + 0.014363297261297703, + 0.010571780614554882, + 0.009531231597065926, + 0.008146429434418678, + 0.00997899193316698, + 0.010583863593637943, + 0.012404467910528183, + 0.007591910660266876, + 0.007023808546364307, + 0.006990598514676094, + 0.00725584989413619, + 0.0067069195210933685, + 0.007082618772983551, + 0.00946823786944151, + 0.01460996177047491, + 0.005888001993298531, + 0.00807322096079588, + 0.01258319802582264, + 0.01053272932767868, + 0.009071249514818192, + 0.008144453167915344, + 0.011236208491027355, + 0.007975204847753048, + 0.014442301355302334, + 0.013283057138323784, + 0.0216058399528265, + 0.01558268815279007, + 0.009145645424723625, + 0.013190657831728458, + 0.014610265381634235, + 0.012854432687163353, + 0.006929073017090559, + 0.008101247251033783, + 0.007262051105499268, + 0.01003294624388218, + 0.008871170692145824, + 0.00830863881856203, + 0.013642598874866962, + 0.011457113549113274, + 0.0037275415379554033, + 0.004555493593215942, + 0.004468842875212431, + 0.006271532736718655, + 0.006650046445429325, + 0.0046468256041407585, + 0.010536500252783298, + 0.005832597613334656, + 0.0031541979406028986, + 0.0031133934389799833, + 0.0043249232694506645, + 0.0035684865433722734, + 0.005288789980113506, + 0.003991980105638504, + 0.007955022156238556, + 0.004915996920317411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.2813659906387329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038099367171525955 + ], + "opposite_logprob_tokens": [ + -3.2880992889404297 + ], + "logprob_reward": -0.038099367171525955, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038099367171525955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11555653810501099, + 0.0722091794013977, + 0.036735136061906815, + 0.08539468795061111, + 0.14060786366462708, + 0.1035698875784874, + 0.20973621308803558, + 0.18047873675823212, + 0.01772300712764263, + 0.037988822907209396 + ], + "credit_map": [ + 0.011580717749893665, + 0.019058918580412865, + 0.01657944917678833, + 0.012225012294948101, + 0.012397811748087406, + 0.012321485206484795, + 0.014453932642936707, + 0.016939209774136543, + 0.008280348032712936, + 0.011459432542324066, + 0.011897817254066467, + 0.008044448681175709, + 0.0067451512441039085, + 0.008229962550103664, + 0.007651898544281721, + 0.009900123812258244, + 0.004713233094662428, + 0.003699058899655938, + 0.00560357142239809, + 0.004065702203661203, + 0.004445253871381283, + 0.004159153904765844, + 0.005254598334431648, + 0.0047945622354745865, + 0.015749182552099228, + 0.009790156967937946, + 0.007056776434183121, + 0.010685674846172333, + 0.00940697081387043, + 0.008927794173359871, + 0.008977096527814865, + 0.014801031909883022, + 0.016662875190377235, + 0.01738339103758335, + 0.012621883302927017, + 0.015662239864468575, + 0.016918973997235298, + 0.015598705969750881, + 0.0193049143999815, + 0.026454873383045197, + 0.013441052287817001, + 0.01408490352332592, + 0.015762263908982277, + 0.0097016841173172, + 0.010413667187094688, + 0.010298934765160084, + 0.011600867845118046, + 0.01826651580631733, + 0.01815483160316944, + 0.017391851171851158, + 0.027419807389378548, + 0.0308589618653059, + 0.021605273708701134, + 0.021540693938732147, + 0.04003365710377693, + 0.03273114189505577, + 0.017919614911079407, + 0.017880983650684357, + 0.026236694306135178, + 0.026557717472314835, + 0.020261051133275032, + 0.019704638049006462, + 0.026970932260155678, + 0.02494710311293602, + 0.0017655518604442477, + 0.0021300073713064194, + 0.0023781575728207827, + 0.002301796106621623, + 0.0020724947098642588, + 0.0017356349853798747, + 0.0028061610646545887, + 0.002533204387873411, + 0.003921928815543652, + 0.0039833830669522285, + 0.004588887095451355, + 0.00472021009773016, + 0.0041309138759970665, + 0.0041915797628462315, + 0.005176982842385769, + 0.007274937815964222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.038099367171525955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera held still, completely directing the viewer's attention to a vibrant moment of nature. A small sparrow perched securely on a rough, mossy apple branch. Its fluffy feathers took on a rich gradation in the light, gradating from a deep brown at the top to a pale grayish-white at the belly. Its small, obsidian-bright eyes fixed intently on the object before them: a large, plump, red apple, its skin smooth as wax, an alluring crimson with hints of yellow-green. Warm afternoon sunlight filtered through the lush green leaves overhead, casting vibrant, fragmented light spots on the bird's back, the curved surface of the apple, and the mottled branches. The sparrow's movements are imbued with rhythm. Quickly and precisely, it lowers its head, its sharp beak firmly pecking at the apple. Each peck lifts up a tiny bit of glistening flesh, revealing the juicy, glistening white flesh at the tiny opening. The entire scene is enveloped in a tranquil and focused atmosphere, dominated by warm golden sunlight, vibrant red, and vibrant green, creating a vibrant, serene, and beautiful pastoral scene.", + "seed": 6992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_15/index_4269_seed_6992_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": -0.3402378689497709, + "vlm_logprob": -0.3402378689497709, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.3402378689497709, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47742873430252075 + ], + "opposite_logprob_tokens": [ + -0.9774287343025208 + ], + "logprob_reward": -0.47742873430252075, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47742873430252075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24547655880451202, + 0.061975207179784775, + 0.06574925780296326, + 0.06706567108631134, + 0.13110876083374023, + 0.09901050478219986, + 0.051223184913396835, + 0.14234140515327454, + 0.07597032189369202, + 0.0600791797041893 + ], + "credit_map": [ + 0.031660787761211395, + 0.019510794430971146, + 0.03454988822340965, + 0.024441564455628395, + 0.02448147162795067, + 0.033362578600645065, + 0.01868318021297455, + 0.05878626927733421, + 0.006117681041359901, + 0.004224397242069244, + 0.00834435224533081, + 0.00770735926926136, + 0.004976872354745865, + 0.008917286060750484, + 0.005619175732135773, + 0.016068076714873314, + 0.005763207096606493, + 0.008308895863592625, + 0.008512189611792564, + 0.011341668665409088, + 0.007540294900536537, + 0.007356713525950909, + 0.008222522214055061, + 0.008703758008778095, + 0.007096817251294851, + 0.008534026332199574, + 0.008862452581524849, + 0.009291843511164188, + 0.00959416851401329, + 0.007715931162238121, + 0.007892550900578499, + 0.008077874779701233, + 0.011424233205616474, + 0.015526458621025085, + 0.01939290016889572, + 0.01702350750565529, + 0.022287555038928986, + 0.015605146996676922, + 0.01371203362941742, + 0.01613691821694374, + 0.010080293752253056, + 0.008291334845125675, + 0.01592038944363594, + 0.012295420281589031, + 0.013938643969595432, + 0.01391727477312088, + 0.006223898846656084, + 0.01834324188530445, + 0.00546244764700532, + 0.0063861398957669735, + 0.008554871194064617, + 0.00585322268307209, + 0.007033232599496841, + 0.006274901330471039, + 0.005590036511421204, + 0.006068331655114889, + 0.014178140088915825, + 0.011345312930643559, + 0.016109686344861984, + 0.017547452822327614, + 0.02596842125058174, + 0.0175265334546566, + 0.013634930364787579, + 0.026030926033854485, + 0.009545748122036457, + 0.006786608602851629, + 0.007227403577417135, + 0.0062181237153708935, + 0.005922709591686726, + 0.012667346745729446, + 0.00665567722171545, + 0.020946692675352097, + 0.008307439275085926, + 0.00667900126427412, + 0.008873356506228447, + 0.008518217131495476, + 0.005251980386674404, + 0.0069726030342280865, + 0.005384490825235844, + 0.010092084296047688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.47742873430252075 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04314572364091873 + ], + "opposite_logprob_tokens": [ + -3.1681456565856934 + ], + "logprob_reward": -0.04314572364091873, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04314572364091873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25248146057128906, + 0.1291223019361496, + 0.04889167845249176, + 0.03854154050350189, + 0.10476680099964142, + 0.10227800160646439, + 0.20416653156280518, + 0.07726328819990158, + 0.019487671554088593, + 0.023000704124569893 + ], + "credit_map": [ + 0.02876356989145279, + 0.014629347249865532, + 0.021603723987936974, + 0.019637716934084892, + 0.014833854511380196, + 0.045500073581933975, + 0.017655296251177788, + 0.08985786885023117, + 0.013314166106283665, + 0.00848825741559267, + 0.01140578743070364, + 0.01177054550498724, + 0.00789710134267807, + 0.021637190133333206, + 0.01168525405228138, + 0.042924005538225174, + 0.006552249193191528, + 0.005379449110478163, + 0.006521143484860659, + 0.006555791944265366, + 0.006854860577732325, + 0.006652973126620054, + 0.004314570687711239, + 0.006060641258955002, + 0.004481988959014416, + 0.004412560723721981, + 0.005161798559129238, + 0.005257978104054928, + 0.006355904508382082, + 0.0046645356342196465, + 0.003613496897742152, + 0.004593275021761656, + 0.01023820135742426, + 0.013448040932416916, + 0.015803957358002663, + 0.013976769521832466, + 0.017080219462513924, + 0.01175561361014843, + 0.011597204022109509, + 0.010866789147257805, + 0.010338632389903069, + 0.008366662077605724, + 0.014356064610183239, + 0.011621423065662384, + 0.01118319109082222, + 0.014568607322871685, + 0.007467185612767935, + 0.02437623217701912, + 0.01572760008275509, + 0.020622890442609787, + 0.022926267236471176, + 0.020312443375587463, + 0.031221993267536163, + 0.02474127896130085, + 0.01770240254700184, + 0.05091164633631706, + 0.006050449796020985, + 0.006758580915629864, + 0.009456600062549114, + 0.009180043824017048, + 0.011200686916708946, + 0.009309439919888973, + 0.008246178738772869, + 0.01706130988895893, + 0.0017065501306205988, + 0.002398253418505192, + 0.0028125757817178965, + 0.003090290119871497, + 0.0025407904759049416, + 0.0022504604421555996, + 0.0024051249492913485, + 0.002283623907715082, + 0.0028911642730236053, + 0.002288751071318984, + 0.002906763693317771, + 0.0035527623258531094, + 0.0033796573989093304, + 0.0024558682925999165, + 0.0025347417686134577, + 0.0029909953009337187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04314572364091873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097875464707613 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097875464707613, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097875464707613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23221825063228607, + 0.12108039855957031, + 0.057992447167634964, + 0.08709583431482315, + 0.03210410475730896, + 0.04491717740893364, + 0.11776570975780487, + 0.047903191298246384, + 0.15012651681900024, + 0.10879629105329514 + ], + "credit_map": [ + 0.016578100621700287, + 0.01607130467891693, + 0.024027260020375252, + 0.029314382001757622, + 0.022445566952228546, + 0.03433478996157646, + 0.023061523213982582, + 0.06638534367084503, + 0.013131550513207912, + 0.010506847873330116, + 0.012498196214437485, + 0.015529346652328968, + 0.010567746125161648, + 0.016847455874085426, + 0.014437812380492687, + 0.027561448514461517, + 0.007211712654680014, + 0.005016421899199486, + 0.00891187135130167, + 0.010256120935082436, + 0.006416697055101395, + 0.007096875924617052, + 0.006457059644162655, + 0.00662568723782897, + 0.009019141085445881, + 0.008565274067223072, + 0.01033092476427555, + 0.010146443732082844, + 0.014233848080039024, + 0.01033804751932621, + 0.007764413021504879, + 0.01669774390757084, + 0.0030152115505188704, + 0.0034330205526202917, + 0.003576934803277254, + 0.003748873481526971, + 0.005190719850361347, + 0.0037775039672851562, + 0.005180809646844864, + 0.004181030672043562, + 0.0033320619259029627, + 0.004194988403469324, + 0.005041005089879036, + 0.004965643398463726, + 0.007990718819200993, + 0.005553757771849632, + 0.006373703945428133, + 0.007465297356247902, + 0.008517023175954819, + 0.010798626579344273, + 0.012533106841146946, + 0.0122819310054183, + 0.01656442880630493, + 0.014541766606271267, + 0.010966879315674305, + 0.03156193718314171, + 0.0053257327526807785, + 0.004800600931048393, + 0.005639626644551754, + 0.005484315101057291, + 0.008538505993783474, + 0.005725892260670662, + 0.005504840984940529, + 0.006883679889142513, + 0.017543887719511986, + 0.016374973580241203, + 0.01708061993122101, + 0.013436342589557171, + 0.0125962495803833, + 0.021204592660069466, + 0.01273121777921915, + 0.039158642292022705, + 0.012545416131615639, + 0.01241794228553772, + 0.01619046926498413, + 0.012687456794083118, + 0.015234650112688541, + 0.013353695161640644, + 0.011391157284379005, + 0.014975498430430889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0097875464707613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.1371098756790161 + ], + "opposite_logprob_tokens": [ + -0.3871099054813385 + ], + "logprob_reward": -1.1371098756790161, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371098756790161, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24583697319030762, + 0.07934517413377762, + 0.08163342624902725, + 0.07669180631637573, + 0.05205901339650154, + 0.0728924348950386, + 0.05614431947469711, + 0.07661125808954239, + 0.22354143857955933, + 0.03524412959814072 + ], + "credit_map": [ + 0.02723156288266182, + 0.020330796018242836, + 0.029062198475003242, + 0.02438371814787388, + 0.020607272163033485, + 0.034713197499513626, + 0.029567858204245567, + 0.059940364211797714, + 0.00847611017525196, + 0.008433797396719456, + 0.0085329944267869, + 0.010249931365251541, + 0.009441999718546867, + 0.009543281979858875, + 0.01140573900192976, + 0.013261321932077408, + 0.00979715958237648, + 0.007447354961186647, + 0.009260308928787708, + 0.012125205248594284, + 0.010228392668068409, + 0.010025225579738617, + 0.010128817521035671, + 0.012620960362255573, + 0.007711329497396946, + 0.007221787702292204, + 0.009379147551953793, + 0.008674711920320988, + 0.009927860461175442, + 0.009455184452235699, + 0.009171846322715282, + 0.015149935148656368, + 0.004527317360043526, + 0.00601199921220541, + 0.006096771452575922, + 0.006402870640158653, + 0.007353462278842926, + 0.006630279123783112, + 0.005443783942610025, + 0.009592530317604542, + 0.008451965637505054, + 0.006337327416986227, + 0.0076601142063736916, + 0.007069411687552929, + 0.009701552800834179, + 0.00896942988038063, + 0.010634415782988071, + 0.01406821794807911, + 0.004616682883352041, + 0.00531473383307457, + 0.0049277725629508495, + 0.006915189325809479, + 0.008255453780293465, + 0.006142067722976208, + 0.011507906951010227, + 0.008464515209197998, + 0.00751522509381175, + 0.006049048155546188, + 0.0072850752621889114, + 0.008382930420339108, + 0.01466471329331398, + 0.008883565664291382, + 0.01162443496286869, + 0.012206265702843666, + 0.027248147875070572, + 0.01863587275147438, + 0.01962958835065365, + 0.0166129469871521, + 0.013545659370720387, + 0.03627542406320572, + 0.018883654847741127, + 0.07271015644073486, + 0.0035650015342980623, + 0.003667109180241823, + 0.004596423823386431, + 0.005018972791731358, + 0.003753471188247204, + 0.004528438672423363, + 0.004008037503808737, + 0.006106674205511808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -1.1371098756790161 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371746465563774 + ], + "opposite_logprob_tokens": [ + -3.408717393875122 + ], + "logprob_reward": -0.03371746465563774, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371746465563774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10575895011425018, + 0.10527439415454865, + 0.03685341775417328, + 0.0450960211455822, + 0.04409124329686165, + 0.036223411560058594, + 0.1729019433259964, + 0.14539939165115356, + 0.05509905517101288, + 0.2533021569252014 + ], + "credit_map": [ + 0.012453380040824413, + 0.009968689642846584, + 0.015072261914610863, + 0.014287035912275314, + 0.008635098114609718, + 0.015717940405011177, + 0.01055067963898182, + 0.019073858857154846, + 0.01379711925983429, + 0.010143652558326721, + 0.009275136515498161, + 0.009664546698331833, + 0.009343421086668968, + 0.01667666807770729, + 0.0074521442875266075, + 0.0289217047393322, + 0.003217492252588272, + 0.0038280177395790815, + 0.00421378668397665, + 0.0050611356273293495, + 0.004106897860765457, + 0.0046893758699297905, + 0.0028945188969373703, + 0.00884218979626894, + 0.004207202233374119, + 0.00549966748803854, + 0.006527291610836983, + 0.006707340944558382, + 0.006672408897429705, + 0.005493570119142532, + 0.004550221376121044, + 0.0054383184760808945, + 0.004611678421497345, + 0.004171822220087051, + 0.005334083456546068, + 0.004481435753405094, + 0.006180928088724613, + 0.005545558873564005, + 0.004834526684135199, + 0.008931208401918411, + 0.0027679046615958214, + 0.003616394940763712, + 0.005654378328472376, + 0.004098459612578154, + 0.006447934545576572, + 0.004722847603261471, + 0.004806547425687313, + 0.004108944442123175, + 0.013477307744324207, + 0.016295213252305984, + 0.01999494805932045, + 0.018661826848983765, + 0.028587838634848595, + 0.020447755232453346, + 0.015808237716555595, + 0.039628807455301285, + 0.01289202831685543, + 0.01129559613764286, + 0.014362473972141743, + 0.016290104016661644, + 0.02720438875257969, + 0.01870352029800415, + 0.013149241916835308, + 0.03150204196572304, + 0.006026767194271088, + 0.006295723374933004, + 0.007404707837849855, + 0.0058136689476668835, + 0.005798622965812683, + 0.007918457500636578, + 0.00537224393337965, + 0.010468860156834126, + 0.030335869640111923, + 0.026449507102370262, + 0.03613459691405296, + 0.028934339061379433, + 0.031194692477583885, + 0.034497104585170746, + 0.021238800138235092, + 0.04451723396778107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.03371746465563774 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera held still, completely directing the viewer's attention to a vibrant moment of nature. A small sparrow perched securely on a rough, mossy apple branch. Its fluffy feathers took on a rich gradation in the light, gradating from a deep brown at the top to a pale grayish-white at the belly. Its small, obsidian-bright eyes fixed intently on the object before them: a large, plump, red apple, its skin smooth as wax, an alluring crimson with hints of yellow-green. Warm afternoon sunlight filtered through the lush green leaves overhead, casting vibrant, fragmented light spots on the bird's back, the curved surface of the apple, and the mottled branches. The sparrow's movements are imbued with rhythm. Quickly and precisely, it lowers its head, its sharp beak firmly pecking at the apple. Each peck lifts up a tiny bit of glistening flesh, revealing the juicy, glistening white flesh at the tiny opening. The entire scene is enveloped in a tranquil and focused atmosphere, dominated by warm golden sunlight, vibrant red, and vibrant green, creating a vibrant, serene, and beautiful pastoral scene.", + "seed": 6993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_15/index_4269_seed_6993_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": -0.16188193149864674, + "vlm_logprob": -0.16188193149864674, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.16188193149864674, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.207720085978508 + ], + "opposite_logprob_tokens": [ + -1.7077200412750244 + ], + "logprob_reward": -0.207720085978508, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.207720085978508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5014893412590027, + 0.11545565724372864, + 0.07283823192119598, + 0.06812586635351181, + 0.03304784372448921, + 0.029341183602809906, + 0.1059378832578659, + 0.024263057857751846, + 0.021491993218660355, + 0.028008926659822464 + ], + "credit_map": [ + 0.033912356942892075, + 0.0657656341791153, + 0.07391434907913208, + 0.04192475974559784, + 0.10653325915336609, + 0.04271615669131279, + 0.09239745885133743, + 0.04432537034153938, + 0.008437806740403175, + 0.007837781682610512, + 0.018998216837644577, + 0.011413428001105785, + 0.014521656557917595, + 0.021479716524481773, + 0.014231751672923565, + 0.018535299226641655, + 0.0069460272789001465, + 0.007591672241687775, + 0.012120585888624191, + 0.006002004723995924, + 0.0072805676609277725, + 0.011297352612018585, + 0.010628572665154934, + 0.010971451178193092, + 0.006087993737310171, + 0.0076552690006792545, + 0.007814591750502586, + 0.006391807924956083, + 0.007424197159707546, + 0.009590806439518929, + 0.013368204236030579, + 0.009792997501790524, + 0.006432496011257172, + 0.003333879867568612, + 0.004415122326463461, + 0.0027091577649116516, + 0.003881930373609066, + 0.0035320324823260307, + 0.0036870460025966167, + 0.005056179128587246, + 0.002482436830177903, + 0.0020655652042478323, + 0.0033283268567174673, + 0.003944073803722858, + 0.005303494166582823, + 0.002368237590417266, + 0.006190046668052673, + 0.003659004345536232, + 0.010146980173885822, + 0.009192965924739838, + 0.012221409007906914, + 0.010753180831670761, + 0.023451974615454674, + 0.014350448735058308, + 0.014191302470862865, + 0.01162962056696415, + 0.001844199956394732, + 0.0020674343686550856, + 0.002922176383435726, + 0.0028774640522897243, + 0.0035911700688302517, + 0.0025634155608713627, + 0.004997273441404104, + 0.0033999255392700434, + 0.002005143091082573, + 0.0022376568522304296, + 0.002535090548917651, + 0.002436745446175337, + 0.0030737416818737984, + 0.0026639439165592194, + 0.004201257135719061, + 0.0023384152445942163, + 0.0035688162315636873, + 0.0031732143834233284, + 0.0038394033908843994, + 0.0025217749644070864, + 0.005044079851359129, + 0.0038124904967844486, + 0.0034367432817816734, + 0.0026124026626348495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.207720085978508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033757686614990234 + ], + "opposite_logprob_tokens": [ + -3.4087576866149902 + ], + "logprob_reward": -0.033757686614990234, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033757686614990234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33903831243515015, + 0.21153181791305542, + 0.06709220260381699, + 0.06680631637573242, + 0.02913147211074829, + 0.0301680751144886, + 0.12318862974643707, + 0.040830228477716446, + 0.0441659577190876, + 0.0480470173060894 + ], + "credit_map": [ + 0.024783963337540627, + 0.038544077426195145, + 0.03995058313012123, + 0.03141480311751366, + 0.06994803249835968, + 0.027228109538555145, + 0.07369548082351685, + 0.03347325697541237, + 0.016496092081069946, + 0.015712972730398178, + 0.03278401866555214, + 0.02177930250763893, + 0.029684221372008324, + 0.03591248393058777, + 0.025967486202716827, + 0.03319524601101875, + 0.006351653952151537, + 0.006463469471782446, + 0.010586913675069809, + 0.006248805671930313, + 0.007201044820249081, + 0.010066491551697254, + 0.01164452638477087, + 0.008529292419552803, + 0.005921042989939451, + 0.006017856299877167, + 0.007507797796279192, + 0.005689585581421852, + 0.006779567804187536, + 0.009430360049009323, + 0.014178394339978695, + 0.011281711049377918, + 0.0031978399492800236, + 0.0030637176241725683, + 0.004121555481106043, + 0.0031040431931614876, + 0.003872587578371167, + 0.003327571554109454, + 0.0033564225304871798, + 0.005087734665721655, + 0.0024842654820531607, + 0.0022991986479610205, + 0.0034501610789448023, + 0.00277505605481565, + 0.005601438693702221, + 0.006109915673732758, + 0.003548372769728303, + 0.003899669973179698, + 0.012050486169755459, + 0.011039885692298412, + 0.013250967487692833, + 0.013014787808060646, + 0.03033836930990219, + 0.012357867322862148, + 0.017169654369354248, + 0.013966611586511135, + 0.004711609333753586, + 0.004087384790182114, + 0.005568237975239754, + 0.004592242185026407, + 0.005753706209361553, + 0.004639645107090473, + 0.0061289225704967976, + 0.0053484803065657616, + 0.004633327946066856, + 0.005036404822021723, + 0.006238152272999287, + 0.004528853110969067, + 0.005638121627271175, + 0.006620028987526894, + 0.006729761604219675, + 0.004741308279335499, + 0.008237115107476711, + 0.005571702029556036, + 0.007431323640048504, + 0.003926715813577175, + 0.00623522512614727, + 0.006570291239768267, + 0.004696932155638933, + 0.005377710331231356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.033757686614990234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192220479249954 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192220479249954, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018192220479249954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21394579112529755, + 0.12784428894519806, + 0.08451788127422333, + 0.05308084562420845, + 0.05857809633016586, + 0.04825965315103531, + 0.0842769593000412, + 0.24828889966011047, + 0.04531823471188545, + 0.035889387130737305 + ], + "credit_map": [ + 0.020246654748916626, + 0.027047619223594666, + 0.032486796379089355, + 0.019058771431446075, + 0.03737855330109596, + 0.023891327902674675, + 0.027126692235469818, + 0.026709377765655518, + 0.015868976712226868, + 0.011483940295875072, + 0.023336615413427353, + 0.012666482478380203, + 0.015409004874527454, + 0.014474875293672085, + 0.017444269731640816, + 0.017160119488835335, + 0.012247373349964619, + 0.008044470101594925, + 0.014353509992361069, + 0.009235108271241188, + 0.008085635490715504, + 0.010230989195406437, + 0.01266736164689064, + 0.009653429500758648, + 0.0037633890751749277, + 0.005720494315028191, + 0.006692578550428152, + 0.005609261337667704, + 0.008356868289411068, + 0.004427737090736628, + 0.011241406202316284, + 0.0072691068053245544, + 0.006048416718840599, + 0.0056428005918860435, + 0.0064393156208097935, + 0.0052072275429964066, + 0.008879797533154488, + 0.007102857809513807, + 0.012058311142027378, + 0.0071993665769696236, + 0.0058110300451517105, + 0.00495014525949955, + 0.005927310325205326, + 0.004977099597454071, + 0.008055122569203377, + 0.004966401495039463, + 0.007002586033195257, + 0.006569954566657543, + 0.008091704919934273, + 0.0068655116483569145, + 0.012372603639960289, + 0.007665466517210007, + 0.01293079275637865, + 0.011315870098769665, + 0.013077677227556705, + 0.011957325041294098, + 0.021693259477615356, + 0.0220863688737154, + 0.025671523064374924, + 0.028955325484275818, + 0.046029508113861084, + 0.029758907854557037, + 0.0486581027507782, + 0.02543589100241661, + 0.004330727271735668, + 0.004530926235020161, + 0.005249872803688049, + 0.005202511325478554, + 0.0059517752379179, + 0.007740271743386984, + 0.006481011398136616, + 0.005831134505569935, + 0.002917086472734809, + 0.0031368625350296497, + 0.004808266647160053, + 0.003808053210377693, + 0.006080520339310169, + 0.004519850015640259, + 0.006885220762342215, + 0.003733523655682802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.018192220479249954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.523391842842102 + ], + "opposite_logprob_tokens": [ + -0.898391842842102 + ], + "logprob_reward": -0.523391842842102, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.523391842842102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28234708309173584, + 0.08746898174285889, + 0.04218028858304024, + 0.10043080896139145, + 0.06711941212415695, + 0.04229357838630676, + 0.06554950028657913, + 0.08572080731391907, + 0.14212045073509216, + 0.08476896584033966 + ], + "credit_map": [ + 0.022747496142983437, + 0.03574852645397186, + 0.03599704056978226, + 0.02586476504802704, + 0.05433334410190582, + 0.022336425259709358, + 0.05646185204386711, + 0.028857680037617683, + 0.011085142381489277, + 0.009192705154418945, + 0.013376133516430855, + 0.009274949319660664, + 0.011750321835279465, + 0.008383318781852722, + 0.01620820350944996, + 0.008198223076760769, + 0.006433189380913973, + 0.003662283532321453, + 0.007577823009341955, + 0.0047928025014698505, + 0.0035785348154604435, + 0.004969631787389517, + 0.0070144436322152615, + 0.004151587840169668, + 0.00856547150760889, + 0.01142895594239235, + 0.012201275676488876, + 0.010184764862060547, + 0.01322801224887371, + 0.013108627870678902, + 0.020468497648835182, + 0.011245214380323887, + 0.0052772583439946175, + 0.006338304374366999, + 0.009987499564886093, + 0.007625441998243332, + 0.009088641963899136, + 0.009168670512735844, + 0.008733177557587624, + 0.010900427587330341, + 0.006815063301473856, + 0.003859669901430607, + 0.006372522562742233, + 0.005894160363823175, + 0.00570075074210763, + 0.0032187963370233774, + 0.005873868707567453, + 0.004558754153549671, + 0.005481029860675335, + 0.007328655570745468, + 0.008306464180350304, + 0.00772477500140667, + 0.0136193186044693, + 0.005541142541915178, + 0.010788675397634506, + 0.0067594521678984165, + 0.009769822470843792, + 0.007739368360489607, + 0.007592205423861742, + 0.010038461536169052, + 0.015523233450949192, + 0.010147703811526299, + 0.01571423001587391, + 0.009195794351398945, + 0.013684995472431183, + 0.015925388783216476, + 0.014297930523753166, + 0.0140618197619915, + 0.020269105210900307, + 0.022492755204439163, + 0.025495877489447594, + 0.015892595052719116, + 0.014165623113512993, + 0.010670578107237816, + 0.011934793554246426, + 0.007046624552458525, + 0.007137606851756573, + 0.013975263573229313, + 0.009606589563190937, + 0.010231900960206985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.523391842842102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347821578383446 + ], + "opposite_logprob_tokens": [ + -3.6513478755950928 + ], + "logprob_reward": -0.026347821578383446, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347821578383446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24606110155582428, + 0.1479347199201584, + 0.0814426839351654, + 0.1316785216331482, + 0.08775297552347183, + 0.057012446224689484, + 0.07697562873363495, + 0.03852054849267006, + 0.0458107627928257, + 0.08681067824363708 + ], + "credit_map": [ + 0.019359160214662552, + 0.040655434131622314, + 0.04758669063448906, + 0.027661381289362907, + 0.04316284507513046, + 0.020366791635751724, + 0.023163313046097755, + 0.02410544827580452, + 0.017802944406867027, + 0.015649432316422462, + 0.01638173870742321, + 0.012953681871294975, + 0.018282486125826836, + 0.021740537136793137, + 0.01972624659538269, + 0.025397637858986855, + 0.009140252135694027, + 0.00667162612080574, + 0.014054601080715656, + 0.007383562624454498, + 0.008095422759652138, + 0.014515426009893417, + 0.008056456223130226, + 0.013525327667593956, + 0.010295975022017956, + 0.014841019175946712, + 0.015514836646616459, + 0.01450842060148716, + 0.018744954839348793, + 0.0183063093572855, + 0.02457704395055771, + 0.01488996297121048, + 0.011238429695367813, + 0.01033200416713953, + 0.017372651025652885, + 0.007571726571768522, + 0.006673427298665047, + 0.012321464717388153, + 0.012509179301559925, + 0.009734083898365498, + 0.005529019515961409, + 0.00456510903313756, + 0.008259223774075508, + 0.007076816633343697, + 0.007822477258741856, + 0.01101592741906643, + 0.006381251383572817, + 0.006362617015838623, + 0.006713042501360178, + 0.006760842632502317, + 0.008241680450737476, + 0.009091987274587154, + 0.016519656404852867, + 0.009133031591773033, + 0.011824510991573334, + 0.008690868504345417, + 0.0032251127995550632, + 0.004319500643759966, + 0.004345891065895557, + 0.004829904064536095, + 0.006808913312852383, + 0.0048146010376513, + 0.00554299633949995, + 0.004633625037968159, + 0.004112363327294588, + 0.005952024832367897, + 0.005461983848363161, + 0.004308691713958979, + 0.006894857622683048, + 0.0064454167149960995, + 0.007275147829204798, + 0.005360272713005543, + 0.009413449093699455, + 0.008489962667226791, + 0.014136006124317646, + 0.008266489021480083, + 0.010242059826850891, + 0.011879702098667622, + 0.0129143251106143, + 0.011468678712844849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.026347821578383446 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera held still, completely directing the viewer's attention to a vibrant moment of nature. A small sparrow perched securely on a rough, mossy apple branch. Its fluffy feathers took on a rich gradation in the light, gradating from a deep brown at the top to a pale grayish-white at the belly. Its small, obsidian-bright eyes fixed intently on the object before them: a large, plump, red apple, its skin smooth as wax, an alluring crimson with hints of yellow-green. Warm afternoon sunlight filtered through the lush green leaves overhead, casting vibrant, fragmented light spots on the bird's back, the curved surface of the apple, and the mottled branches. The sparrow's movements are imbued with rhythm. Quickly and precisely, it lowers its head, its sharp beak firmly pecking at the apple. Each peck lifts up a tiny bit of glistening flesh, revealing the juicy, glistening white flesh at the tiny opening. The entire scene is enveloped in a tranquil and focused atmosphere, dominated by warm golden sunlight, vibrant red, and vibrant green, creating a vibrant, serene, and beautiful pastoral scene.", + "seed": 6994, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_15/index_4269_seed_6994_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": -0.21067161969840525, + "vlm_logprob": -0.21067161969840525, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.21067161969840525, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4366746246814728 + ], + "opposite_logprob_tokens": [ + -1.0616745948791504 + ], + "logprob_reward": -0.4366746246814728, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4366746246814728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41917139291763306, + 0.026611564680933952, + 0.1293339729309082, + 0.05753851309418678, + 0.10723607987165451, + 0.12281481176614761, + 0.057505231350660324, + 0.02496291697025299, + 0.024869859218597412, + 0.029955709353089333 + ], + "credit_map": [ + 0.03827294334769249, + 0.04915937781333923, + 0.03647148981690407, + 0.03613825887441635, + 0.11009924858808517, + 0.05273826792836189, + 0.06402712315320969, + 0.032264627516269684, + 0.0031298974063247442, + 0.00550705473870039, + 0.0036040290724486113, + 0.0023063370026648045, + 0.0036478450056165457, + 0.0029616267420351505, + 0.003005403559654951, + 0.0024493695236742496, + 0.012897149659693241, + 0.019648894667625427, + 0.010623357258737087, + 0.016528695821762085, + 0.025618432089686394, + 0.017051460221409798, + 0.015659648925065994, + 0.011306321248412132, + 0.006136883515864611, + 0.008852174505591393, + 0.007555828429758549, + 0.007869618944823742, + 0.009847949258983135, + 0.006717933341860771, + 0.006480067037045956, + 0.004078053869307041, + 0.01041327603161335, + 0.014493614435195923, + 0.016323084011673927, + 0.013895178213715553, + 0.01566394791007042, + 0.012334796600043774, + 0.012836605310440063, + 0.011275574564933777, + 0.010375372134149075, + 0.021372931078076363, + 0.0114291375502944, + 0.012751953676342964, + 0.026971042156219482, + 0.01777164451777935, + 0.0157266054302454, + 0.006416119635105133, + 0.005275642964988947, + 0.008542348630726337, + 0.006593334022909403, + 0.01052767876535654, + 0.0061013344675302505, + 0.006769971922039986, + 0.009641820564866066, + 0.00405309721827507, + 0.004123125225305557, + 0.002746290061622858, + 0.004167478997260332, + 0.004953551571816206, + 0.002034299774095416, + 0.00280768983066082, + 0.002017810707911849, + 0.0021126705687493086, + 0.0026423707604408264, + 0.005295715760439634, + 0.003685937263071537, + 0.002723309677094221, + 0.0021782114636152983, + 0.003337372560054064, + 0.002783393021672964, + 0.0022235456854104996, + 0.0032350525725632906, + 0.004776095971465111, + 0.0027787236031144857, + 0.0049499752931296825, + 0.00342369731515646, + 0.004725464154034853, + 0.003763314802199602, + 0.0023033844772726297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.4366746246814728 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04874254763126373 + ], + "opposite_logprob_tokens": [ + -3.0487425327301025 + ], + "logprob_reward": -0.04874254763126373, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04874254763126373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4467427432537079, + 0.046132221817970276, + 0.11964183300733566, + 0.04792094603180885, + 0.09868964552879333, + 0.08752162009477615, + 0.08015388995409012, + 0.021377883851528168, + 0.013559123501181602, + 0.03826006501913071 + ], + "credit_map": [ + 0.040947142988443375, + 0.05579867213964462, + 0.03730718046426773, + 0.03841930255293846, + 0.11317934095859528, + 0.057992372661828995, + 0.06948437541723251, + 0.03361434116959572, + 0.0044854083098471165, + 0.006629788316786289, + 0.004404623992741108, + 0.004514262080192566, + 0.010576730594038963, + 0.0060130455531179905, + 0.006598498672246933, + 0.0029098670929670334, + 0.012318838387727737, + 0.020495587959885597, + 0.010195798240602016, + 0.015341591089963913, + 0.020948218181729317, + 0.016147900372743607, + 0.013148137368261814, + 0.011045756749808788, + 0.004811875522136688, + 0.008826833218336105, + 0.006383135449141264, + 0.006191229447722435, + 0.007950537838041782, + 0.005967156030237675, + 0.004920481238514185, + 0.002869699615985155, + 0.009390313178300858, + 0.014675687067210674, + 0.014693988487124443, + 0.012708951719105244, + 0.014218417927622795, + 0.011812503449618816, + 0.011447898112237453, + 0.009741878136992455, + 0.00785132683813572, + 0.015890948474407196, + 0.007136158645153046, + 0.007715736515820026, + 0.019179943948984146, + 0.0131631875410676, + 0.012083877809345722, + 0.004500444047152996, + 0.007974973879754543, + 0.009739230386912823, + 0.009886323474347591, + 0.018902167677879333, + 0.0067648873664438725, + 0.009749597869813442, + 0.009667221456766129, + 0.007469487376511097, + 0.0025796866975724697, + 0.002831923309713602, + 0.0033702277578413486, + 0.002483739983290434, + 0.002718512900173664, + 0.0025103853549808264, + 0.003274607239291072, + 0.0016088002594187856, + 0.0017070575850084424, + 0.001696392660960555, + 0.001672819023951888, + 0.0020444311667233706, + 0.0019229083554819226, + 0.0016828239895403385, + 0.0017851517768576741, + 0.0010475394083186984, + 0.003958175890147686, + 0.006740811746567488, + 0.0036966828629374504, + 0.006518525071442127, + 0.004416388925164938, + 0.005873103626072407, + 0.004041345790028572, + 0.003015031572431326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.04874254763126373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199829384684563 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018199829384684563, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199829384684563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3198866546154022, + 0.03245294839143753, + 0.17718522250652313, + 0.0508657768368721, + 0.06451462209224701, + 0.07934273034334183, + 0.1743316352367401, + 0.02236672304570675, + 0.03881781920790672, + 0.04023582860827446 + ], + "credit_map": [ + 0.028787894174456596, + 0.041559066623449326, + 0.025349562987685204, + 0.028534874320030212, + 0.07732122391462326, + 0.04373718053102493, + 0.05049216374754906, + 0.024104714393615723, + 0.003443067194893956, + 0.0049893176183104515, + 0.004130768124014139, + 0.00349896727129817, + 0.004440756980329752, + 0.004204845987260342, + 0.004939444828778505, + 0.002805780852213502, + 0.017676357179880142, + 0.02671315334737301, + 0.017441879957914352, + 0.02574644424021244, + 0.029765509068965912, + 0.02289639785885811, + 0.020481852814555168, + 0.016463618725538254, + 0.0075360615737736225, + 0.005307849962264299, + 0.00560373067855835, + 0.007898645475506783, + 0.0071831573732197285, + 0.005492519121617079, + 0.0072603230364620686, + 0.004583490546792746, + 0.0068155089393258095, + 0.009302731603384018, + 0.007924668490886688, + 0.00831815879791975, + 0.009929995983839035, + 0.007198148872703314, + 0.008951161056756973, + 0.006074253935366869, + 0.006649585906416178, + 0.01355714350938797, + 0.007061010226607323, + 0.007742969319224358, + 0.016176141798496246, + 0.011507819406688213, + 0.012446333654224873, + 0.004201726987957954, + 0.018397362902760506, + 0.02792387269437313, + 0.019732113927602768, + 0.03589499741792679, + 0.015408857725560665, + 0.0214006919413805, + 0.016049271449446678, + 0.019524475559592247, + 0.002557390136644244, + 0.003720264881849289, + 0.0019689046312123537, + 0.002405949868261814, + 0.003571171313524246, + 0.0029276579152792692, + 0.0036199335008859634, + 0.0015954503323882818, + 0.003598153591156006, + 0.006714011076837778, + 0.00530980434268713, + 0.008432737551629543, + 0.004037648905068636, + 0.00475015863776207, + 0.0033400028478354216, + 0.002635302022099495, + 0.0039250534027814865, + 0.006701115984469652, + 0.0035247416235506535, + 0.0067696827463805676, + 0.005737488158047199, + 0.0060316938906908035, + 0.004479177296161652, + 0.0030668778344988823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.018199829384684563 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233866572380066 + ], + "opposite_logprob_tokens": [ + -0.8983866572380066 + ], + "logprob_reward": -0.5233866572380066, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233866572380066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5028703212738037, + 0.04269038513302803, + 0.051846105605363846, + 0.052948806434869766, + 0.04239611700177193, + 0.08781884610652924, + 0.06389026343822479, + 0.05031632259488106, + 0.07167968899011612, + 0.03354315087199211 + ], + "credit_map": [ + 0.0462593138217926, + 0.06717802584171295, + 0.04001317173242569, + 0.04242505133152008, + 0.12361720949411392, + 0.06740676611661911, + 0.07970359921455383, + 0.036267153918743134, + 0.00426940293982625, + 0.006270761135965586, + 0.00400037644430995, + 0.005174946505576372, + 0.00873671192675829, + 0.006063086912035942, + 0.005296872928738594, + 0.0028782282024621964, + 0.005756175145506859, + 0.00866716355085373, + 0.0035508135333657265, + 0.0062246196903288364, + 0.010702759027481079, + 0.006509473547339439, + 0.005990307312458754, + 0.004444791469722986, + 0.0064478605054318905, + 0.005187669303268194, + 0.007078249007463455, + 0.0070630344562232494, + 0.009377745911478996, + 0.00571790337562561, + 0.007031591143459082, + 0.005044755060225725, + 0.004447077866643667, + 0.005327264312654734, + 0.004504481330513954, + 0.0047095404006540775, + 0.007012170739471912, + 0.004962509963661432, + 0.007132279220968485, + 0.0043007927015423775, + 0.007448404096066952, + 0.01437043584883213, + 0.00884541030973196, + 0.009468183852732182, + 0.01807910017669201, + 0.012361199595034122, + 0.01228080503642559, + 0.004965307656675577, + 0.0057791112922132015, + 0.012963631190359592, + 0.0051412442699074745, + 0.006713743321597576, + 0.008698373101651669, + 0.008342512883245945, + 0.011803007684648037, + 0.004448637831956148, + 0.009578217752277851, + 0.005174129270017147, + 0.008388830348849297, + 0.007742891553789377, + 0.004453758243471384, + 0.005619114264845848, + 0.004886338021606207, + 0.004473046399652958, + 0.007616496179252863, + 0.007378495763987303, + 0.01366097666323185, + 0.017193200066685677, + 0.005213343072682619, + 0.008324231952428818, + 0.006231670267879963, + 0.006061268970370293, + 0.0034080848563462496, + 0.005173993296921253, + 0.0031380411237478256, + 0.005290172062814236, + 0.004361474886536598, + 0.004917622543871403, + 0.004625595174729824, + 0.002628164365887642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.5233866572380066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a small sparrow perched on a rough, mossy apple branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sparrow have fluffy feathers that gradate from deep brown on top to pale grayish-white on the belly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a large, plump red apple with smooth, wax-like skin and hints of yellow-green visible in front of the sparrow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sparrow quickly and precisely lower its head to peck at the apple, revealing glistening white flesh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still at a level angle, with warm afternoon sunlight filtering through green leaves to cast fragmented light spots?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354439556598663 + ], + "opposite_logprob_tokens": [ + -3.6513545513153076 + ], + "logprob_reward": -0.026354439556598663, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354439556598663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29533037543296814, + 0.03154624253511429, + 0.13374295830726624, + 0.06327749043703079, + 0.0542818158864975, + 0.08532439917325974, + 0.06317432969808578, + 0.1025678887963295, + 0.12856923043727875, + 0.04218525066971779 + ], + "credit_map": [ + 0.02712382562458515, + 0.03581751883029938, + 0.026600101962685585, + 0.02704000286757946, + 0.07514638453722, + 0.039228156208992004, + 0.040833573788404465, + 0.02354082092642784, + 0.0030702161602675915, + 0.005669113714247942, + 0.005415694788098335, + 0.0030535829719156027, + 0.004961814731359482, + 0.0037346850149333477, + 0.003088810248300433, + 0.002552325138822198, + 0.0132099948823452, + 0.019452180713415146, + 0.012150594964623451, + 0.017266487702727318, + 0.026823624968528748, + 0.01712510734796524, + 0.015473797917366028, + 0.012241176329553127, + 0.0076569910161197186, + 0.011853238567709923, + 0.009151278994977474, + 0.00876824650913477, + 0.007519267965108156, + 0.007936418987810612, + 0.005910669453442097, + 0.004481376614421606, + 0.005703616421669722, + 0.008930623531341553, + 0.005300754215568304, + 0.005987533833831549, + 0.010219870135188103, + 0.00675986148416996, + 0.0069083282724022865, + 0.004471223335713148, + 0.0076941694132983685, + 0.015326826833188534, + 0.006480774376541376, + 0.008124646730720997, + 0.018807461485266685, + 0.013004605658352375, + 0.011799849569797516, + 0.00408606743440032, + 0.005881850142031908, + 0.006954588927328587, + 0.008699674159288406, + 0.013811955228447914, + 0.006619203370064497, + 0.006880906410515308, + 0.008551569655537605, + 0.0057745762169361115, + 0.016120417043566704, + 0.009648852050304413, + 0.019790073856711388, + 0.022086896002292633, + 0.007834300398826599, + 0.011359572410583496, + 0.007834565825760365, + 0.007893210276961327, + 0.013684254139661789, + 0.01449588593095541, + 0.024129532277584076, + 0.029342295601963997, + 0.010878502391278744, + 0.015732983127236366, + 0.01074395515024662, + 0.009561818093061447, + 0.004548933357000351, + 0.005940599832683802, + 0.0044050635769963264, + 0.007832091301679611, + 0.004709289874881506, + 0.006892160512506962, + 0.004807173274457455, + 0.003049938241019845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.026354439556598663 + } + ], + "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/0000146/group_6/index_tensor([5776, 5776])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_tensor([5776, 5776])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..814fbf9f1378d324bfb96bd4001831b270841bd0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_tensor([5776, 5776])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, low-angle shot slowly pans left, sweeping across a winding section of ancient city wall, coiling like a blue-gray dragon across the rolling hills. It's dusk, and the lingering rays of the setting sun paint the sky a warm, melancholic orange-red, illuminating the wall's weathered masonry, covered in cracks and covered in tenacious moss. Suddenly, without warning, a section of the wall in the center of the frame begins to tremble violently, then with a resounding, muffled explosion, it bursts outward and crumbles. Thousands of bricks and stones, driven by gravity, cascade down like a waterfall, colliding with each other and shattering into countless fragments of stone and powder. A huge and dense yellow-brown dust cloud suddenly rose into the sky, quickly engulfing the collapsed walls and surrounding vegetation. Under the backlight of the setting sun, the edge of the rolling dust was outlined with a magnificent golden outline, and the light penetrated through it, forming visible beams of light. The whole scene was full of a sense of destructive power and a tragic epic atmosphere.", + "seed": 12892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_5776_seed_12892_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": -0.019014590373262762, + "vlm_logprob": -0.019014590373262762, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.019014590373262762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.061531659215688705 + ], + "opposite_logprob_tokens": [ + -2.9365315437316895 + ], + "logprob_reward": -0.061531659215688705, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.061531659215688705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32130250334739685, + 0.06663815677165985, + 0.04741490259766579, + 0.050657719373703, + 0.05206012353301048, + 0.04856216907501221, + 0.054648254066705704, + 0.18753331899642944, + 0.09867508709430695, + 0.0725078359246254 + ], + "credit_map": [ + 0.019232744351029396, + 0.015688326209783554, + 0.04410776495933533, + 0.051195770502090454, + 0.04439639672636986, + 0.05246533453464508, + 0.05054290592670441, + 0.043673254549503326, + 0.002912204945459962, + 0.004270592704415321, + 0.0060350787825882435, + 0.01020083762705326, + 0.008573927916586399, + 0.010762804187834263, + 0.01426460687071085, + 0.009618104435503483, + 0.0029485574923455715, + 0.0032793425489217043, + 0.004933228250592947, + 0.005787343252450228, + 0.00452661095187068, + 0.006860140711069107, + 0.012452338822185993, + 0.006627338007092476, + 0.002453758381307125, + 0.00413070572540164, + 0.006556965410709381, + 0.006418296135962009, + 0.0032674050889909267, + 0.0043385224416852, + 0.012634910643100739, + 0.01085715088993311, + 0.003956229891628027, + 0.0046536801382899284, + 0.005701462738215923, + 0.005791992414742708, + 0.0044638304971158504, + 0.007944011129438877, + 0.009139922447502613, + 0.010408989153802395, + 0.005194160621613264, + 0.003845532890409231, + 0.0069019366055727005, + 0.004811418708413839, + 0.00336013943888247, + 0.005737566854804754, + 0.008958603255450726, + 0.00975281372666359, + 0.0037000083830207586, + 0.003161367727443576, + 0.004426565486937761, + 0.004768054932355881, + 0.0049244496040046215, + 0.009677965193986893, + 0.011855854652822018, + 0.012133982963860035, + 0.01227755006402731, + 0.011699952185153961, + 0.019495295360684395, + 0.018722984939813614, + 0.023041781038045883, + 0.034242741763591766, + 0.03686048835515976, + 0.031192542985081673, + 0.006566015537828207, + 0.006525460630655289, + 0.01323776040226221, + 0.018771085888147354, + 0.01160791702568531, + 0.016236403957009315, + 0.012478112243115902, + 0.013252331875264645, + 0.005802478175610304, + 0.005027382634580135, + 0.006712328642606735, + 0.008266729302704334, + 0.010650490410625935, + 0.016029616817831993, + 0.010709055699408054, + 0.009309754706919193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.061531659215688705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553482316434383 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553482316434383, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553482316434383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4082716703414917, + 0.16001421213150024, + 0.06791504472494125, + 0.055025044828653336, + 0.020975446328520775, + 0.03553805872797966, + 0.05851098150014877, + 0.08309731632471085, + 0.06816090643405914, + 0.0424913726747036 + ], + "credit_map": [ + 0.016915954649448395, + 0.013545216992497444, + 0.03786875680088997, + 0.058835841715335846, + 0.03658583015203476, + 0.04140514135360718, + 0.11023704707622528, + 0.09287787973880768, + 0.007232727017253637, + 0.009630179964005947, + 0.02106589265167713, + 0.03418324515223503, + 0.01398039236664772, + 0.020039884373545647, + 0.034515149891376495, + 0.01936674304306507, + 0.004378882236778736, + 0.0035541043616831303, + 0.00894131325185299, + 0.0064412690699100494, + 0.004926483146846294, + 0.011217565275728703, + 0.01907493732869625, + 0.009380490519106388, + 0.0032885666005313396, + 0.0034579234197735786, + 0.007110966369509697, + 0.004798373207449913, + 0.0037022861652076244, + 0.004093233961611986, + 0.014806799590587616, + 0.013766894116997719, + 0.0019415150163695216, + 0.0021419664844870567, + 0.0026799379847943783, + 0.00339136621914804, + 0.0019524620147421956, + 0.002842722227796912, + 0.0030338838696479797, + 0.002991593675687909, + 0.0041788602247834206, + 0.005072885192930698, + 0.004265739582479, + 0.003557164454832673, + 0.0031121952924877405, + 0.0041744341142475605, + 0.00592059176415205, + 0.005256188567727804, + 0.007442332338541746, + 0.006826888769865036, + 0.005910047795623541, + 0.005364326760172844, + 0.006657533813267946, + 0.009444596245884895, + 0.0077444263733923435, + 0.00912083126604557, + 0.005805318709462881, + 0.005386352073401213, + 0.007902740500867367, + 0.009172756224870682, + 0.01024240255355835, + 0.015174665488302708, + 0.016443749889731407, + 0.012969334609806538, + 0.006648891139775515, + 0.00536789046600461, + 0.008274788968265057, + 0.013119535520672798, + 0.007667373865842819, + 0.010260875336825848, + 0.00968177616596222, + 0.00713977450504899, + 0.002864438109099865, + 0.002519097877666354, + 0.0028283472638577223, + 0.0059400456957519054, + 0.004593894816935062, + 0.00911910180002451, + 0.008112666197121143, + 0.006513781379908323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012553482316434383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3089814782142639, + 0.1680014282464981, + 0.031768374145030975, + 0.060672637075185776, + 0.07189405709505081, + 0.091557577252388, + 0.0691257119178772, + 0.06812389940023422, + 0.10025271773338318, + 0.029622133821249008 + ], + "credit_map": [ + 0.023900533095002174, + 0.014969783835113049, + 0.04600740224123001, + 0.05271773040294647, + 0.02386191487312317, + 0.03835609555244446, + 0.05844500660896301, + 0.05072298273444176, + 0.012193548493087292, + 0.010279728099703789, + 0.014957178384065628, + 0.021193163469433784, + 0.016852304339408875, + 0.026409825310111046, + 0.04149596020579338, + 0.024619709700345993, + 0.0025799376890063286, + 0.002995080780237913, + 0.004026465583592653, + 0.004056042991578579, + 0.0023424222599714994, + 0.0037994266021996737, + 0.006283336319029331, + 0.005685660522431135, + 0.0026871433947235346, + 0.0029473991598933935, + 0.0073611801490187645, + 0.007397529669106007, + 0.0028250799514353275, + 0.005088870879262686, + 0.018008915707468987, + 0.014356515370309353, + 0.00550014479085803, + 0.00682422099635005, + 0.007914179004728794, + 0.008789153769612312, + 0.00612023938447237, + 0.0071194469928741455, + 0.01828436367213726, + 0.011342305690050125, + 0.00998341292142868, + 0.014111802913248539, + 0.01083951722830534, + 0.010850094258785248, + 0.009909999556839466, + 0.01161242090165615, + 0.012475762516260147, + 0.011774556711316109, + 0.008087337017059326, + 0.006507414393126965, + 0.00694249477237463, + 0.008757153525948524, + 0.009779457934200764, + 0.010402406565845013, + 0.009575152769684792, + 0.009074297733604908, + 0.009630125015974045, + 0.008239340968430042, + 0.008599681779742241, + 0.00806141085922718, + 0.007822182960808277, + 0.009274912998080254, + 0.008394826203584671, + 0.008101418614387512, + 0.007017964497208595, + 0.006879543885588646, + 0.012483716942369938, + 0.02118569053709507, + 0.009378984570503235, + 0.014030830934643745, + 0.01502004824578762, + 0.014255933463573456, + 0.003533147508278489, + 0.003514382755383849, + 0.0029028533026576042, + 0.0031139308121055365, + 0.0046076299622654915, + 0.0046770693734288216, + 0.0034111959394067526, + 0.0038619227707386017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545360252261162 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545360252261162, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545360252261162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10007862001657486, + 0.03972350060939789, + 0.08744664490222931, + 0.03022151067852974, + 0.04743356257677078, + 0.10067529231309891, + 0.17421594262123108, + 0.23588836193084717, + 0.08830352127552032, + 0.09601298719644547 + ], + "credit_map": [ + 0.004445632919669151, + 0.0027645276859402657, + 0.011970357038080692, + 0.014573423191905022, + 0.0048142122104763985, + 0.01085853110998869, + 0.026664400473237038, + 0.023987529799342155, + 0.0022776885889470577, + 0.0018726199632510543, + 0.007270380388945341, + 0.0055957334116101265, + 0.003166568698361516, + 0.005641009658575058, + 0.008864900097250938, + 0.005034600384533405, + 0.005668358411639929, + 0.006466307211667299, + 0.010136408731341362, + 0.010284597054123878, + 0.006281232461333275, + 0.01152848917990923, + 0.023389386013150215, + 0.013691864907741547, + 0.0019138521747663617, + 0.0028610744047909975, + 0.004454887472093105, + 0.00453979941084981, + 0.002253351965919137, + 0.0030319183133542538, + 0.006479897536337376, + 0.00468673137947917, + 0.0036031415220350027, + 0.005253694951534271, + 0.004985481034964323, + 0.006007058080285788, + 0.005042300559580326, + 0.007416530977934599, + 0.008315096609294415, + 0.00681025767698884, + 0.008490841835737228, + 0.011432833038270473, + 0.012393280863761902, + 0.011333348229527473, + 0.010705542750656605, + 0.011910726316273212, + 0.016405915841460228, + 0.018002808094024658, + 0.020689750090241432, + 0.016725575551390648, + 0.014180372469127178, + 0.017121311277151108, + 0.01342326682060957, + 0.020293479785323143, + 0.036746539175510406, + 0.035035643726587296, + 0.015864627435803413, + 0.015643123537302017, + 0.025080259889364243, + 0.02862047217786312, + 0.02900964394211769, + 0.03838305175304413, + 0.04902861267328262, + 0.03425857424736023, + 0.00566769577562809, + 0.0061936331912875175, + 0.013633593916893005, + 0.016390493139624596, + 0.009513013996183872, + 0.008658521808683872, + 0.01807750202715397, + 0.010169061832129955, + 0.007828451693058014, + 0.006879448890686035, + 0.010304737836122513, + 0.010543600656092167, + 0.013046961277723312, + 0.02152581699192524, + 0.013172872364521027, + 0.012711099348962307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.012545360252261162 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249644163995981 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249644163995981, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249644163995981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08116038143634796, + 0.07084139436483383, + 0.08592838793992996, + 0.07482530176639557, + 0.04485129937529564, + 0.06164538487792015, + 0.07690440118312836, + 0.09998191148042679, + 0.356819212436676, + 0.047042276710271835 + ], + "credit_map": [ + 0.006138174328953028, + 0.0033744049724191427, + 0.011200899258255959, + 0.014099537394940853, + 0.0060225557535886765, + 0.01325712725520134, + 0.015766168013215065, + 0.011301515623927116, + 0.004233078565448523, + 0.0036945887841284275, + 0.007599545642733574, + 0.011378823779523373, + 0.007051602005958557, + 0.00902445800602436, + 0.01824401132762432, + 0.009615285322070122, + 0.004946725908666849, + 0.006043575704097748, + 0.0099259614944458, + 0.010431160219013691, + 0.006786453537642956, + 0.012945391237735748, + 0.022511616349220276, + 0.012337501160800457, + 0.0033331250306218863, + 0.0039242771454155445, + 0.010247108526527882, + 0.007951862178742886, + 0.004893176257610321, + 0.00609264662489295, + 0.022541526705026627, + 0.015841586515307426, + 0.004211085848510265, + 0.003576297778636217, + 0.005071770399808884, + 0.005596743896603584, + 0.004057046491652727, + 0.0051435017958283424, + 0.010811549611389637, + 0.006383304949849844, + 0.006216447334736586, + 0.008933437056839466, + 0.006474888883531094, + 0.006423684768378735, + 0.00783593486994505, + 0.006503966636955738, + 0.010050182230770588, + 0.009206840768456459, + 0.007152222096920013, + 0.008191796019673347, + 0.008857311680912971, + 0.007840891368687153, + 0.006424819119274616, + 0.012444308958947659, + 0.012705420143902302, + 0.013287636451423168, + 0.008591359481215477, + 0.007793117314577103, + 0.017451953142881393, + 0.011055960319936275, + 0.01058103796094656, + 0.012579197064042091, + 0.017138943076133728, + 0.014790345914661884, + 0.024482514709234238, + 0.02465236373245716, + 0.04633539915084839, + 0.0727924108505249, + 0.035336434841156006, + 0.05755520612001419, + 0.043285809457302094, + 0.052379049360752106, + 0.0034642419777810574, + 0.003055237466469407, + 0.0035383407957851887, + 0.004433673806488514, + 0.005358542315661907, + 0.012355961836874485, + 0.006856399122625589, + 0.007979879155755043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.005249644163995981 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, low-angle shot slowly pans left, sweeping across a winding section of ancient city wall, coiling like a blue-gray dragon across the rolling hills. It's dusk, and the lingering rays of the setting sun paint the sky a warm, melancholic orange-red, illuminating the wall's weathered masonry, covered in cracks and covered in tenacious moss. Suddenly, without warning, a section of the wall in the center of the frame begins to tremble violently, then with a resounding, muffled explosion, it bursts outward and crumbles. Thousands of bricks and stones, driven by gravity, cascade down like a waterfall, colliding with each other and shattering into countless fragments of stone and powder. A huge and dense yellow-brown dust cloud suddenly rose into the sky, quickly engulfing the collapsed walls and surrounding vegetation. Under the backlight of the setting sun, the edge of the rolling dust was outlined with a magnificent golden outline, and the light penetrated through it, forming visible beams of light. The whole scene was full of a sense of destructive power and a tragic epic atmosphere.", + "seed": 12893, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_5776_seed_12893_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": -0.014631472807377578, + "vlm_logprob": -0.014631472807377578, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.014631472807377578, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044662076979875565 + ], + "opposite_logprob_tokens": [ + -3.4196619987487793 + ], + "logprob_reward": -0.044662076979875565, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.044662076979875565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059922292828559875, + 0.1854502558708191, + 0.5499527454376221, + 0.03963366895914078, + 0.03788107633590698, + 0.014380711130797863, + 0.03311176970601082, + 0.025730442255735397, + 0.02160976640880108, + 0.032327260822057724 + ], + "credit_map": [ + 0.006992443930357695, + 0.0079812565818429, + 0.010485293343663216, + 0.009006500244140625, + 0.005817723460495472, + 0.00651452736929059, + 0.006569600664079189, + 0.006554947234690189, + 0.008644175715744495, + 0.010101143270730972, + 0.060933392494916916, + 0.07285735011100769, + 0.004765673074871302, + 0.006600537803024054, + 0.012641888111829758, + 0.008906112983822823, + 0.019337743520736694, + 0.011606043204665184, + 0.19744306802749634, + 0.0511128306388855, + 0.014316948130726814, + 0.009200329892337322, + 0.20776335895061493, + 0.03917241469025612, + 0.003442117478698492, + 0.005009142681956291, + 0.007107059936970472, + 0.003946215845644474, + 0.00478846812620759, + 0.0032857779879122972, + 0.007166644092649221, + 0.004888243041932583, + 0.003901412943378091, + 0.0040724496357142925, + 0.004781636875122786, + 0.004492154344916344, + 0.0051251002587378025, + 0.002888928633183241, + 0.006471545901149511, + 0.006147849839180708, + 0.0018383113201707602, + 0.0019699044059962034, + 0.001855345326475799, + 0.0017553097568452358, + 0.0018256379989907146, + 0.0015687678242102265, + 0.0018489890499040484, + 0.0017184453317895532, + 0.002563831629231572, + 0.0034221091773360968, + 0.0038731165695935488, + 0.0031164800748229027, + 0.0034775424282997847, + 0.006105632521212101, + 0.006784386001527309, + 0.00376866920851171, + 0.002376699820160866, + 0.002634188160300255, + 0.0030361604876816273, + 0.0042002820409834385, + 0.002575257793068886, + 0.0054666404612362385, + 0.0035590173210948706, + 0.0018821974517777562, + 0.0023148360196501017, + 0.002937064040452242, + 0.0022047595120966434, + 0.0014568574260920286, + 0.002139553427696228, + 0.004643159918487072, + 0.003116234904155135, + 0.002797300461679697, + 0.0030111095402389765, + 0.0037872791290283203, + 0.006866913288831711, + 0.0020080239046365023, + 0.0075002978555858135, + 0.0027166290674358606, + 0.0030483328737318516, + 0.0033886765595525503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.044662076979875565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562546879053116 + ], + "opposite_logprob_tokens": [ + -4.3875627517700195 + ], + "logprob_reward": -0.012562546879053116, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562546879053116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16756735742092133, + 0.16021795570850372, + 0.39555948972702026, + 0.02276814915239811, + 0.16521961987018585, + 0.014988778159022331, + 0.014825215563178062, + 0.015670493245124817, + 0.01327562890946865, + 0.029907360672950745 + ], + "credit_map": [ + 0.016064057126641273, + 0.017038613557815552, + 0.027460908517241478, + 0.023987283930182457, + 0.014384089969098568, + 0.025108911097049713, + 0.021085688844323158, + 0.022437801584601402, + 0.0080541567876935, + 0.009816877543926239, + 0.05144792050123215, + 0.057452473789453506, + 0.006711337249726057, + 0.006909032352268696, + 0.010125512257218361, + 0.009700629860162735, + 0.013929875567555428, + 0.008886425755918026, + 0.14141444861888885, + 0.038286443799734116, + 0.01088047306984663, + 0.008331789635121822, + 0.14683498442173004, + 0.026994982734322548, + 0.0029912865720689297, + 0.0026899126823991537, + 0.0036070961505174637, + 0.0023269751109182835, + 0.001809685374610126, + 0.002461186842992902, + 0.0030068333726376295, + 0.003875171532854438, + 0.015561276115477085, + 0.027771472930908203, + 0.02130495011806488, + 0.014964863657951355, + 0.03108186461031437, + 0.013313841074705124, + 0.017131736502051353, + 0.024089600890874863, + 0.0015570473624393344, + 0.0023893872275948524, + 0.0016282470896840096, + 0.00152731838170439, + 0.0023871525190770626, + 0.0016813452821224928, + 0.0021005338057875633, + 0.001717744511552155, + 0.0009776619262993336, + 0.00219652708619833, + 0.0018147471128031611, + 0.0013219993561506271, + 0.0016072026919573545, + 0.0022875547874718904, + 0.002484044758602977, + 0.0021354760974645615, + 0.0014986301539465785, + 0.0023955409415066242, + 0.0028695494402199984, + 0.001747544971294701, + 0.001956589287146926, + 0.002175590256229043, + 0.001808645436540246, + 0.0012184014776721597, + 0.001416432554833591, + 0.001378107932396233, + 0.0015160454204306006, + 0.001075529376976192, + 0.0016449891263619065, + 0.0013592534232884645, + 0.002551834797486663, + 0.0023334354627877474, + 0.0034352820366621017, + 0.005140035878866911, + 0.004717391449958086, + 0.00221699895337224, + 0.0035166051238775253, + 0.0033242213539779186, + 0.0035687063354998827, + 0.0039881146512925625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.012562546879053116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641822263598442 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641822263598442, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641822263598442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07481693476438522, + 0.1921118199825287, + 0.13443009555339813, + 0.031283020973205566, + 0.17993591725826263, + 0.11145607382059097, + 0.04659133777022362, + 0.04579104855656624, + 0.02691265568137169, + 0.15667107701301575 + ], + "credit_map": [ + 0.004585131071507931, + 0.007470368407666683, + 0.011266644112765789, + 0.008971082046627998, + 0.008854354731738567, + 0.011500289663672447, + 0.009635519236326218, + 0.012533544562757015, + 0.010170099325478077, + 0.015291422605514526, + 0.05779800936579704, + 0.0669349804520607, + 0.006400018464773893, + 0.006433346774429083, + 0.014293435961008072, + 0.014790511690080166, + 0.006419377867132425, + 0.005655358079820871, + 0.050592899322509766, + 0.01606805995106697, + 0.0036207062657922506, + 0.0037089278921484947, + 0.040269263088703156, + 0.008095511235296726, + 0.0029811959248036146, + 0.004571526311337948, + 0.003973295912146568, + 0.005161198787391186, + 0.003916027490049601, + 0.0034584924578666687, + 0.00433339411392808, + 0.002887888578698039, + 0.01564677432179451, + 0.028903262689709663, + 0.02301921881735325, + 0.017479615285992622, + 0.03302228823304176, + 0.017343439161777496, + 0.019254790619015694, + 0.025266535580158234, + 0.010625575669109821, + 0.016787797212600708, + 0.011474397033452988, + 0.00949514377862215, + 0.01843206025660038, + 0.016918648034334183, + 0.01440033596009016, + 0.013322112150490284, + 0.004052341915667057, + 0.0037057758308947086, + 0.004360332153737545, + 0.0033123937901109457, + 0.0060375588946044445, + 0.014386670663952827, + 0.005025259219110012, + 0.005711004137992859, + 0.004271922167390585, + 0.007867182604968548, + 0.0037448948714882135, + 0.0024842594284564257, + 0.007951444946229458, + 0.010914118029177189, + 0.004347692709416151, + 0.004209531936794519, + 0.0020922906696796417, + 0.002753814682364464, + 0.00302133453078568, + 0.0018755362834781408, + 0.002339816652238369, + 0.0036125669721513987, + 0.007036528550088406, + 0.00418076803907752, + 0.014797641895711422, + 0.022323738783597946, + 0.03792956471443176, + 0.00940697267651558, + 0.040050145238637924, + 0.011293326504528522, + 0.00949796847999096, + 0.011371737346053123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.008641822263598442 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099300131201744 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099300131201744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099300131201744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062394119799137115, + 0.40593740344047546, + 0.2657594382762909, + 0.0626802071928978, + 0.053613726049661636, + 0.061221618205308914, + 0.029341956600546837, + 0.015557053498923779, + 0.017835790291428566, + 0.025658663362264633 + ], + "credit_map": [ + 0.005688142962753773, + 0.008546827360987663, + 0.008739879354834557, + 0.00786121841520071, + 0.007953259162604809, + 0.009748915210366249, + 0.00647006556391716, + 0.0073858099058270454, + 0.015823977068066597, + 0.016437608748674393, + 0.15008926391601562, + 0.1662745177745819, + 0.007589342538267374, + 0.01177237369120121, + 0.024037567898631096, + 0.013912774622440338, + 0.00956519041210413, + 0.007856078445911407, + 0.07899198681116104, + 0.02406812086701393, + 0.009400337003171444, + 0.007374512497335672, + 0.10256657004356384, + 0.025936642661690712, + 0.007020460441708565, + 0.008099619299173355, + 0.009730431251227856, + 0.007713510654866695, + 0.006682686973363161, + 0.0042729866690933704, + 0.010584978386759758, + 0.008575532585382462, + 0.0054451217874884605, + 0.008405860513448715, + 0.006206707563251257, + 0.00516689196228981, + 0.009645096957683563, + 0.005409610457718372, + 0.006065417546778917, + 0.007269017864018679, + 0.0068304045125842094, + 0.009297949261963367, + 0.0076787336729466915, + 0.0044120727106928825, + 0.009363751858472824, + 0.00967888068407774, + 0.006584396585822105, + 0.007375427521765232, + 0.0024868508335202932, + 0.002190752886235714, + 0.0026914903428405523, + 0.001893489621579647, + 0.00317199993878603, + 0.008807304315268993, + 0.004387738183140755, + 0.0037123295478522778, + 0.0013996672350913286, + 0.0023236735723912716, + 0.0017414567992091179, + 0.001343312906101346, + 0.002119053155183792, + 0.00383083987981081, + 0.001468951115384698, + 0.001330097671598196, + 0.0016209681052714586, + 0.0016269623301923275, + 0.0021413543727248907, + 0.0010951502481475472, + 0.0016442174091935158, + 0.0023685661144554615, + 0.004516188055276871, + 0.0028223826084285975, + 0.002894425531849265, + 0.002942498540505767, + 0.0060734557919204235, + 0.0018334490014240146, + 0.006054618861526251, + 0.0017086940351873636, + 0.001961066387593746, + 0.0021904539316892624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.004099300131201744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031916177831590176 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031916177831590176, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031916177831590176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024586504325270653, + 0.18840859830379486, + 0.4759375751018524, + 0.03350627422332764, + 0.141561359167099, + 0.012718284502625465, + 0.06153499707579613, + 0.02832207642495632, + 0.009579413570463657, + 0.023844925686717033 + ], + "credit_map": [ + 0.002478339010849595, + 0.00305322022177279, + 0.0036838070955127478, + 0.0033356351777911186, + 0.0021070935763418674, + 0.0033847405575215816, + 0.003147316863760352, + 0.003396350424736738, + 0.006805095821619034, + 0.006427326705306768, + 0.07193595170974731, + 0.08046509325504303, + 0.0022449688985943794, + 0.003957443870604038, + 0.01033481489866972, + 0.006237909197807312, + 0.016730714589357376, + 0.00947884563356638, + 0.16635707020759583, + 0.04180685058236122, + 0.012714706361293793, + 0.007204276975244284, + 0.18382401764392853, + 0.037821076810359955, + 0.002951261820271611, + 0.005531581584364176, + 0.003950833342969418, + 0.004327013622969389, + 0.004382225219160318, + 0.002814915496855974, + 0.00507681630551815, + 0.0044716233387589455, + 0.012869649566709995, + 0.02356797270476818, + 0.017231767997145653, + 0.013296691700816154, + 0.027768483385443687, + 0.012403038330376148, + 0.01434293296188116, + 0.020080823451280594, + 0.001457505626603961, + 0.001751619391143322, + 0.001615728484466672, + 0.0019443848868831992, + 0.0010989431757479906, + 0.0011990333441644907, + 0.002197026275098324, + 0.001454043434932828, + 0.005156556144356728, + 0.0075320093892514706, + 0.005844517145305872, + 0.005252127069979906, + 0.006557091139256954, + 0.01365911029279232, + 0.009404845535755157, + 0.008128739893436432, + 0.0029918416403234005, + 0.003496229648590088, + 0.002156774280592799, + 0.0020591667853295803, + 0.004621345549821854, + 0.0058310371823608875, + 0.004644659347832203, + 0.0025210243184119463, + 0.0011618477292358875, + 0.0011777360923588276, + 0.0017440284136682749, + 0.0007819675956852734, + 0.0011015809141099453, + 0.0014848245773464441, + 0.0010868305107578635, + 0.0010405973298475146, + 0.002185878111049533, + 0.0028874825220555067, + 0.005796912591904402, + 0.0014949956675991416, + 0.006478143390268087, + 0.0018797670491039753, + 0.001632924540899694, + 0.0014888232108205557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.0031916177831590176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, low-angle shot slowly pans left, sweeping across a winding section of ancient city wall, coiling like a blue-gray dragon across the rolling hills. It's dusk, and the lingering rays of the setting sun paint the sky a warm, melancholic orange-red, illuminating the wall's weathered masonry, covered in cracks and covered in tenacious moss. Suddenly, without warning, a section of the wall in the center of the frame begins to tremble violently, then with a resounding, muffled explosion, it bursts outward and crumbles. Thousands of bricks and stones, driven by gravity, cascade down like a waterfall, colliding with each other and shattering into countless fragments of stone and powder. A huge and dense yellow-brown dust cloud suddenly rose into the sky, quickly engulfing the collapsed walls and surrounding vegetation. Under the backlight of the setting sun, the edge of the rolling dust was outlined with a magnificent golden outline, and the light penetrated through it, forming visible beams of light. The whole scene was full of a sense of destructive power and a tragic epic atmosphere.", + "seed": 12894, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_5776_seed_12894_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": -0.030141874682158232, + "vlm_logprob": -0.030141874682158232, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.030141874682158232, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08730203658342361 + ], + "opposite_logprob_tokens": [ + -2.5873019695281982 + ], + "logprob_reward": -0.08730203658342361, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08730203658342361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17427118122577667, + 0.3151558041572571, + 0.10719433426856995, + 0.12380041927099228, + 0.04591045528650284, + 0.03179984912276268, + 0.07852865010499954, + 0.05174119025468826, + 0.03844968602061272, + 0.033148393034935 + ], + "credit_map": [ + 0.00658827181905508, + 0.004827866796404123, + 0.013037244789302349, + 0.04123971238732338, + 0.006120352074503899, + 0.006792737636715174, + 0.0722343698143959, + 0.023430628702044487, + 0.011750582605600357, + 0.004746874328702688, + 0.020437704399228096, + 0.04940303787589073, + 0.00877522211521864, + 0.03491540253162384, + 0.12794263660907745, + 0.057184383273124695, + 0.004224568605422974, + 0.002868840703740716, + 0.009993089362978935, + 0.010612021200358868, + 0.0028438579756766558, + 0.005821162834763527, + 0.034890066832304, + 0.03594072908163071, + 0.004663546569645405, + 0.002284328918904066, + 0.010669006034731865, + 0.0117250457406044, + 0.004111113026738167, + 0.007888750173151493, + 0.047682516276836395, + 0.03477611392736435, + 0.0034107102546840906, + 0.00367086473852396, + 0.006481898948550224, + 0.011668043211102486, + 0.0035792638082057238, + 0.004615360870957375, + 0.005726906470954418, + 0.006757410708814859, + 0.002043158980086446, + 0.004287128336727619, + 0.0045391167514026165, + 0.006630884483456612, + 0.0018391840858384967, + 0.004230625927448273, + 0.003490663832053542, + 0.004739085678011179, + 0.007219118997454643, + 0.009192763827741146, + 0.010551521554589272, + 0.008249135687947273, + 0.006717702839523554, + 0.01390470564365387, + 0.00660133408382535, + 0.016092367470264435, + 0.0063573336228728294, + 0.010816825553774834, + 0.00525326281785965, + 0.005102272145450115, + 0.005084975156933069, + 0.00661823945119977, + 0.004431252833455801, + 0.008077027276158333, + 0.0031516251619905233, + 0.003952223341912031, + 0.005097655579447746, + 0.006765875965356827, + 0.0045989155769348145, + 0.0028879682067781687, + 0.006039418745785952, + 0.0059560067020356655, + 0.0028656499926000834, + 0.009163775481283665, + 0.002905686618760228, + 0.002565499860793352, + 0.005820068065077066, + 0.0026450231671333313, + 0.003614598885178566, + 0.0035680935252457857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.08730203658342361 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603561773896217 + ], + "opposite_logprob_tokens": [ + -3.895603656768799 + ], + "logprob_reward": -0.020603561773896217, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603561773896217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15525266528129578, + 0.08390727639198303, + 0.4971604347229004, + 0.08981339633464813, + 0.04480787366628647, + 0.02404640056192875, + 0.04130015894770622, + 0.006713604088872671, + 0.016376245766878128, + 0.04062199220061302 + ], + "credit_map": [ + 0.005407322198152542, + 0.0052961851470172405, + 0.016977613791823387, + 0.035169731825590134, + 0.00614881468936801, + 0.00882827490568161, + 0.0509943962097168, + 0.02643030881881714, + 0.004591905511915684, + 0.003466004505753517, + 0.008471023291349411, + 0.010582992807030678, + 0.005051387473940849, + 0.02066454105079174, + 0.019666505977511406, + 0.011412911117076874, + 0.02983871102333069, + 0.014830075204372406, + 0.043453555554151535, + 0.04545888677239418, + 0.01432335376739502, + 0.025236956775188446, + 0.14533741772174835, + 0.17868146300315857, + 0.0031104295048862696, + 0.002520728623494506, + 0.007050078362226486, + 0.009278944693505764, + 0.0024231085553765297, + 0.004742024466395378, + 0.03445679321885109, + 0.02623128332197666, + 0.00328299542888999, + 0.0032517900690436363, + 0.0066851237788796425, + 0.0089136753231287, + 0.0029137686360627413, + 0.005316751543432474, + 0.006749080494046211, + 0.0076946839690208435, + 0.0023203857708722353, + 0.0021983981132507324, + 0.0028140870854258537, + 0.004673258401453495, + 0.0014617847045883536, + 0.0025299282278865576, + 0.003432396799325943, + 0.00461615901440382, + 0.005010019987821579, + 0.005987898912280798, + 0.0053113666363060474, + 0.006534794345498085, + 0.0040472326800227165, + 0.006431235931813717, + 0.003032882232218981, + 0.004944725893437862, + 0.001065440010279417, + 0.0009108172380365431, + 0.0007384096970781684, + 0.0009004990570247173, + 0.0008101562852971256, + 0.0007018426549620926, + 0.0007144059636630118, + 0.0008720327168703079, + 0.0016628794837743044, + 0.001651784754358232, + 0.00166643678676337, + 0.0020885749254375696, + 0.0022411535028368235, + 0.002941207028925419, + 0.0021217919420450926, + 0.002002414781600237, + 0.004113250877708197, + 0.013545393943786621, + 0.0017503194976598024, + 0.003300902433693409, + 0.008287467993795872, + 0.002224745461717248, + 0.003048414597287774, + 0.004351493902504444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.020603561773896217 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811221942305565 + ], + "opposite_logprob_tokens": [ + -3.288112163543701 + ], + "logprob_reward": -0.03811221942305565, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811221942305565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030053751543164253, + 0.016826042905449867, + 0.6247274875640869, + 0.10606807470321655, + 0.0676712617278099, + 0.07264084368944168, + 0.05090242251753807, + 0.006917884107679129, + 0.017121070995926857, + 0.007071126718074083 + ], + "credit_map": [ + 0.0016861041076481342, + 0.001748076407238841, + 0.003454525489360094, + 0.0062306467443704605, + 0.0014954234939068556, + 0.0019780765287578106, + 0.00954520795494318, + 0.003915693145245314, + 0.001330880681052804, + 0.0012151440605521202, + 0.0025172545574605465, + 0.0031984865199774504, + 0.0012547512305900455, + 0.003042200580239296, + 0.0022758180275559425, + 0.001991508062928915, + 0.035096507519483566, + 0.013883627951145172, + 0.047777291387319565, + 0.04302294924855232, + 0.011797633953392506, + 0.028458528220653534, + 0.20369623601436615, + 0.24099478125572205, + 0.003457681043073535, + 0.001123548368923366, + 0.008161096833646297, + 0.009124180302023888, + 0.0025439888704568148, + 0.004854017868638039, + 0.04368886724114418, + 0.03311469405889511, + 0.005479861982166767, + 0.004493519198149443, + 0.009866383858025074, + 0.0169199388474226, + 0.004896603524684906, + 0.007402262184768915, + 0.007764934096485376, + 0.010847765021026134, + 0.007806992623955011, + 0.007862384431064129, + 0.00935243722051382, + 0.011680854484438896, + 0.0043359240517020226, + 0.006437762640416622, + 0.008699807338416576, + 0.016464682295918465, + 0.006319890264421701, + 0.008103135041892529, + 0.005923756863921881, + 0.006748410407453775, + 0.0030735181644558907, + 0.004810438025742769, + 0.005020282696932554, + 0.010902997106313705, + 0.0012727119028568268, + 0.0010119415819644928, + 0.0007126332493498921, + 0.0006854230305179954, + 0.0006191510474309325, + 0.001121225650422275, + 0.0006529971724376082, + 0.0008418008219450712, + 0.0014573457883670926, + 0.0018733474425971508, + 0.0020476251374930143, + 0.003080724272876978, + 0.0020876277703791857, + 0.0013893760042265058, + 0.0028242941480129957, + 0.0023607322946190834, + 0.0006262622191570699, + 0.0006809683400206268, + 0.0008382071973755956, + 0.0007716515683569014, + 0.0010564376134425402, + 0.0009446002077311277, + 0.0012143005151301622, + 0.0009386998135596514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.03811221942305565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249301316216588 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.00249301316216588, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249301316216588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19956253468990326, + 0.14458532631397247, + 0.0979238823056221, + 0.13028383255004883, + 0.030336827039718628, + 0.212917298078537, + 0.040574681013822556, + 0.05053534358739853, + 0.0330876000225544, + 0.06019262969493866 + ], + "credit_map": [ + 0.007348966784775257, + 0.0037614682223647833, + 0.014557207934558392, + 0.05192740261554718, + 0.005432715639472008, + 0.007993623614311218, + 0.08196413516998291, + 0.02657702937722206, + 0.005922236014157534, + 0.0025197237264364958, + 0.0098253870382905, + 0.026161227375268936, + 0.004875310696661472, + 0.012376025319099426, + 0.05841444432735443, + 0.024490980431437492, + 0.0039001672994345427, + 0.002843602793291211, + 0.009433366358280182, + 0.00923070963472128, + 0.003049086546525359, + 0.005429613869637251, + 0.031347718089818954, + 0.03268962353467941, + 0.00464461836963892, + 0.003284916514530778, + 0.009883484803140163, + 0.011174735613167286, + 0.003914295695722103, + 0.005993467289954424, + 0.051215168088674545, + 0.040173161774873734, + 0.0023557774256914854, + 0.0018290788866579533, + 0.0040673986077308655, + 0.005429571028798819, + 0.002284315414726734, + 0.00420298520475626, + 0.0039104376919567585, + 0.006257266737520695, + 0.02036452293395996, + 0.025828829035162926, + 0.030881036072969437, + 0.04277275130152702, + 0.015115086920559406, + 0.021739128977060318, + 0.023713605478405952, + 0.032502368092536926, + 0.004862654488533735, + 0.005879492498934269, + 0.004508280195295811, + 0.005610848776996136, + 0.0036097574047744274, + 0.005996895022690296, + 0.0033394533675163984, + 0.006767303217202425, + 0.009305736050009727, + 0.006568884011358023, + 0.006338874809443951, + 0.009201355278491974, + 0.0040228646248579025, + 0.004278214182704687, + 0.004957092460244894, + 0.0058623263612389565, + 0.002561911940574646, + 0.0034559655468910933, + 0.004710576497018337, + 0.005636309739202261, + 0.0035388425458222628, + 0.0038691184017807245, + 0.004749243147671223, + 0.0045656366273760796, + 0.005271957255899906, + 0.00824619922786951, + 0.006998652126640081, + 0.005688479170203209, + 0.013765226118266582, + 0.0056215617805719376, + 0.0079653887078166, + 0.006635167170315981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.00249301316216588 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198542468249798 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198542468249798, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198542468249798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23393166065216064, + 0.08763502538204193, + 0.37838107347488403, + 0.051501352339982986, + 0.021491171792149544, + 0.07051169127225876, + 0.04979170113801956, + 0.02525467798113823, + 0.02459675446152687, + 0.056904956698417664 + ], + "credit_map": [ + 0.008614873513579369, + 0.0047658635303378105, + 0.017037902027368546, + 0.059402696788311005, + 0.006596183869987726, + 0.009320286102592945, + 0.09684494882822037, + 0.03134886175394058, + 0.0034152669832110405, + 0.001495798584073782, + 0.005765352863818407, + 0.013803818263113499, + 0.0029144673608243465, + 0.01090224552899599, + 0.034907467663288116, + 0.014430593699216843, + 0.02114628441631794, + 0.008285739459097385, + 0.028670182451605797, + 0.024673057720065117, + 0.007021044380962849, + 0.017011601477861404, + 0.12464801967144012, + 0.14692507684230804, + 0.0020502624101936817, + 0.0017411451553925872, + 0.003986121155321598, + 0.004709916654974222, + 0.0016635583015158772, + 0.004399068187922239, + 0.01881025917828083, + 0.014141014777123928, + 0.0012706988491117954, + 0.0011896940413862467, + 0.002792105544358492, + 0.004701307043433189, + 0.001248640357516706, + 0.0021433394867926836, + 0.0030154837295413017, + 0.005129901226609945, + 0.007541946601122618, + 0.012225618585944176, + 0.009770374745130539, + 0.01236334815621376, + 0.00358287850394845, + 0.007128387689590454, + 0.007466116454452276, + 0.010433012619614601, + 0.005031917244195938, + 0.005735031329095364, + 0.0063087837770581245, + 0.006409728433936834, + 0.004351203329861164, + 0.008787309750914574, + 0.0043022120371460915, + 0.008865512907505035, + 0.0036650816909968853, + 0.0029238536953926086, + 0.003177219070494175, + 0.002810706617310643, + 0.0022075732704252005, + 0.004159134346991777, + 0.0031905423384159803, + 0.0031205634586513042, + 0.0019356253324076533, + 0.0023631243966519833, + 0.002458043396472931, + 0.0037498914171010256, + 0.003259611316025257, + 0.003151080571115017, + 0.004415135830640793, + 0.0032642397563904524, + 0.003731095464900136, + 0.005643895361572504, + 0.005754636600613594, + 0.007428235374391079, + 0.007187352981418371, + 0.005504268687218428, + 0.014328001998364925, + 0.0073274653404951096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.002198542468249798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, low-angle shot slowly pans left, sweeping across a winding section of ancient city wall, coiling like a blue-gray dragon across the rolling hills. It's dusk, and the lingering rays of the setting sun paint the sky a warm, melancholic orange-red, illuminating the wall's weathered masonry, covered in cracks and covered in tenacious moss. Suddenly, without warning, a section of the wall in the center of the frame begins to tremble violently, then with a resounding, muffled explosion, it bursts outward and crumbles. Thousands of bricks and stones, driven by gravity, cascade down like a waterfall, colliding with each other and shattering into countless fragments of stone and powder. A huge and dense yellow-brown dust cloud suddenly rose into the sky, quickly engulfing the collapsed walls and surrounding vegetation. Under the backlight of the setting sun, the edge of the rolling dust was outlined with a magnificent golden outline, and the light penetrated through it, forming visible beams of light. The whole scene was full of a sense of destructive power and a tragic epic atmosphere.", + "seed": 12895, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000146/group_6/index_5776_seed_12895_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": -0.041904053278267385, + "vlm_logprob": -0.041904053278267385, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.041904053278267385, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1721303015947342 + ], + "opposite_logprob_tokens": [ + -1.9221303462982178 + ], + "logprob_reward": -0.1721303015947342, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1721303015947342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11354492604732513, + 0.12123698741197586, + 0.06363531202077866, + 0.13239561021327972, + 0.08978596329689026, + 0.057735931128263474, + 0.05431656539440155, + 0.14068444073200226, + 0.14325116574764252, + 0.08341306447982788 + ], + "credit_map": [ + 0.007254682015627623, + 0.009145405143499374, + 0.016993286088109016, + 0.017817575484514236, + 0.008944323286414146, + 0.01640712283551693, + 0.015499567613005638, + 0.021482976153492928, + 0.006712539587169886, + 0.0068015484139323235, + 0.021414542570710182, + 0.02897280640900135, + 0.009799247607588768, + 0.02137802727520466, + 0.013347246684134007, + 0.012811035849153996, + 0.0030428392346948385, + 0.0027124362532049417, + 0.009256194345653057, + 0.0156918466091156, + 0.0039609624072909355, + 0.009324871934950352, + 0.01490536704659462, + 0.004740800708532333, + 0.006999615579843521, + 0.004121917299926281, + 0.04495910927653313, + 0.021005375310778618, + 0.003964162431657314, + 0.0204791072756052, + 0.016874991357326508, + 0.0139913409948349, + 0.008490403182804585, + 0.008608470670878887, + 0.01738503947854042, + 0.010201816447079182, + 0.006842528935521841, + 0.0125297075137496, + 0.012326745316386223, + 0.013401256874203682, + 0.0031867465004324913, + 0.003393553663045168, + 0.00886211171746254, + 0.010470566339790821, + 0.005446387454867363, + 0.00914300512522459, + 0.006997401360422373, + 0.010236162692308426, + 0.004048936069011688, + 0.0033581098541617393, + 0.008519873023033142, + 0.00738501688465476, + 0.005192487966269255, + 0.004277265630662441, + 0.010910485871136189, + 0.010624391958117485, + 0.011150384321808815, + 0.009763546288013458, + 0.021869724616408348, + 0.019795868545770645, + 0.01263574231415987, + 0.01646735891699791, + 0.02507704868912697, + 0.023924779146909714, + 0.016603222116827965, + 0.014899476431310177, + 0.01940774731338024, + 0.01706668548285961, + 0.011726484633982182, + 0.01102658174932003, + 0.029351964592933655, + 0.023169012740254402, + 0.0170319601893425, + 0.017286477610468864, + 0.008933770470321178, + 0.006777061149477959, + 0.005701691843569279, + 0.0060932510532438755, + 0.00905531458556652, + 0.012533538974821568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.1721303015947342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060496248304844 + ], + "opposite_logprob_tokens": [ + -3.8956048488616943 + ], + "logprob_reward": -0.02060496248304844, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060496248304844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17098647356033325, + 0.28061696887016296, + 0.11116369068622589, + 0.10658116638660431, + 0.0406196154654026, + 0.025288129225373268, + 0.05241027846932411, + 0.050384584814310074, + 0.1212095394730568, + 0.0407395102083683 + ], + "credit_map": [ + 0.009587970562279224, + 0.011024776846170425, + 0.028672993183135986, + 0.03584200143814087, + 0.01339449267834425, + 0.0342910997569561, + 0.016873136162757874, + 0.021300025284290314, + 0.01738562248647213, + 0.01918189786374569, + 0.050132639706134796, + 0.053221046924591064, + 0.030608169734477997, + 0.05084262788295746, + 0.028368236497044563, + 0.03087673895061016, + 0.005281729623675346, + 0.006558742839843035, + 0.01839037984609604, + 0.021000493317842484, + 0.010014606639742851, + 0.015929605811834335, + 0.021507393568754196, + 0.012480754405260086, + 0.0065915840677917, + 0.005499531980603933, + 0.020848486572504044, + 0.016698146238923073, + 0.004808439407497644, + 0.012737510725855827, + 0.02099045179784298, + 0.01840701699256897, + 0.00250712176784873, + 0.0029066880233585835, + 0.0066238935105502605, + 0.004967477638274431, + 0.0020155785605311394, + 0.0038511899765580893, + 0.009367079474031925, + 0.008380590938031673, + 0.0012398341204971075, + 0.0015513818943873048, + 0.0040437583811581135, + 0.0045463829301297665, + 0.0021716938354074955, + 0.0031622613314539194, + 0.0038604605942964554, + 0.004712357651442289, + 0.002979602897539735, + 0.0028610369190573692, + 0.010174915194511414, + 0.0070239827036857605, + 0.004318602848798037, + 0.004916395526379347, + 0.01265216339379549, + 0.007483576890081167, + 0.003547192784026265, + 0.003423495916649699, + 0.00833264458924532, + 0.008161927573382854, + 0.005256049335002899, + 0.007488652598112822, + 0.009197136387228966, + 0.004977489355951548, + 0.008863701485097408, + 0.008981598541140556, + 0.013600364327430725, + 0.0131772980093956, + 0.009612794034183025, + 0.008560806512832642, + 0.03221835941076279, + 0.026194624602794647, + 0.0045659905299544334, + 0.004103562328964472, + 0.007494062650948763, + 0.006224960088729858, + 0.003942382987588644, + 0.003205631859600544, + 0.0065876804292202, + 0.004615241661667824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02060496248304844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738083902746439 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738083902746439, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738083902746439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1961124688386917, + 0.14950662851333618, + 0.14919613301753998, + 0.05215344950556755, + 0.06527659296989441, + 0.05128997191786766, + 0.0831025093793869, + 0.08290441334247589, + 0.06888509541749954, + 0.10157284140586853 + ], + "credit_map": [ + 0.012584628537297249, + 0.013611874543130398, + 0.03600030764937401, + 0.04263293743133545, + 0.01126097608357668, + 0.037483930587768555, + 0.019763246178627014, + 0.02277454547584057, + 0.0071295215748250484, + 0.007407783064991236, + 0.028469296172261238, + 0.0416218601167202, + 0.008676734752953053, + 0.02540159784257412, + 0.015772225335240364, + 0.015027591027319431, + 0.007896983064711094, + 0.007115963380783796, + 0.018574101850390434, + 0.034105971455574036, + 0.010619527660310268, + 0.021358726546168327, + 0.0359782911837101, + 0.013546556234359741, + 0.0030029062181711197, + 0.0025649270974099636, + 0.013821407221257687, + 0.009945317171514034, + 0.0021407143212854862, + 0.0065496875904500484, + 0.006960654631257057, + 0.007167832460254431, + 0.0044563207775354385, + 0.004024267662316561, + 0.00878950022161007, + 0.01315219234675169, + 0.0029450226575136185, + 0.005989512894302607, + 0.015132013708353043, + 0.01078775990754366, + 0.00469745509326458, + 0.0035764016211032867, + 0.009045611135661602, + 0.008863630704581738, + 0.004850615747272968, + 0.007335841190069914, + 0.005912307184189558, + 0.007008102722465992, + 0.010290725156664848, + 0.008170385845005512, + 0.014017531648278236, + 0.013493357226252556, + 0.006078748963773251, + 0.008420830592513084, + 0.012713843956589699, + 0.009917078539729118, + 0.007822573184967041, + 0.006735525093972683, + 0.013865331187844276, + 0.011554093100130558, + 0.006015782244503498, + 0.00908928643912077, + 0.015136132948100567, + 0.012685687281191349, + 0.006070340517908335, + 0.006853126920759678, + 0.00828439462929964, + 0.0065850745886564255, + 0.0073503837920725346, + 0.004709798377007246, + 0.01197200920432806, + 0.01705995574593544, + 0.013109473511576653, + 0.010890014469623566, + 0.026020722463726997, + 0.015348055399954319, + 0.006815082859247923, + 0.006603987887501717, + 0.009908412583172321, + 0.012877080589532852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006738083902746439 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595379201695323 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595379201695323, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595379201695323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1253303587436676, + 0.027644861489534378, + 0.05677197128534317, + 0.0720502957701683, + 0.09061826765537262, + 0.061422478407621384, + 0.158201664686203, + 0.1531854271888733, + 0.062240660190582275, + 0.1925339698791504 + ], + "credit_map": [ + 0.006571270991116762, + 0.007059451658278704, + 0.01886172965168953, + 0.027124948799610138, + 0.008649997413158417, + 0.024917203933000565, + 0.014532295987010002, + 0.017613446339964867, + 0.001506916480138898, + 0.0014852348249405622, + 0.004244715441018343, + 0.006243960000574589, + 0.001913192798383534, + 0.004834167193621397, + 0.004299241118133068, + 0.003117434913292527, + 0.002284848364070058, + 0.002257090061903, + 0.006992859300225973, + 0.014936617575585842, + 0.0034981193020939827, + 0.008284371346235275, + 0.013335822150111198, + 0.0051822420209646225, + 0.003616140689700842, + 0.0025426275096833706, + 0.01851520501077175, + 0.008418316952884197, + 0.0023734953720122576, + 0.00777176208794117, + 0.017322968691587448, + 0.011489778757095337, + 0.0038726485799998045, + 0.004638835322111845, + 0.017686428502202034, + 0.00809550005942583, + 0.006192463915795088, + 0.008306358009576797, + 0.017876427620649338, + 0.023949606344103813, + 0.0029157614335417747, + 0.004248632583767176, + 0.010830752551555634, + 0.011551758274435997, + 0.004674274008721113, + 0.00689513236284256, + 0.010568858124315739, + 0.00973731093108654, + 0.009505532681941986, + 0.014048054814338684, + 0.03500078245997429, + 0.020017586648464203, + 0.018571875989437103, + 0.013251925818622112, + 0.026886606588959694, + 0.020919298753142357, + 0.011916480958461761, + 0.009138551540672779, + 0.02828443981707096, + 0.023884305730462074, + 0.01597701944410801, + 0.01286312472075224, + 0.026499267667531967, + 0.02462223544716835, + 0.005729670636355877, + 0.00645403889939189, + 0.006160675082355738, + 0.008837311528623104, + 0.005129085388034582, + 0.0054199048317968845, + 0.015352213755249977, + 0.009157765656709671, + 0.019863052293658257, + 0.016828959807753563, + 0.026068907231092453, + 0.024536199867725372, + 0.01950846053659916, + 0.02103465422987938, + 0.041627123951911926, + 0.02306661382317543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.00595379201695323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 low-angle shot that pans slowly to the left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ancient city wall blue-gray in color and coiling across rolling hills?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky orange-red due to the setting sun at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a section of the wall in the center tremble, explode, and crumble into a dust cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the dust cloud yellow-brown with a golden outline and visible light beams penetrating it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093126393854618 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093126393854618, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093126393854618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09974205493927002, + 0.04129266366362572, + 0.05015690624713898, + 0.08201626688241959, + 0.10423226654529572, + 0.0272886510938406, + 0.0628066286444664, + 0.06828037649393082, + 0.09773671627044678, + 0.36644747853279114 + ], + "credit_map": [ + 0.005379974842071533, + 0.007009343709796667, + 0.016885779798030853, + 0.022898081690073013, + 0.006086685229092836, + 0.01711617223918438, + 0.012352243065834045, + 0.01201377809047699, + 0.0019025037763640285, + 0.0026751835830509663, + 0.0060133724473416805, + 0.0107529042288661, + 0.0028897812590003014, + 0.006700425408780575, + 0.00608394481241703, + 0.00427455035969615, + 0.0026226516347378492, + 0.0027127235662192106, + 0.0068810018710792065, + 0.009120658040046692, + 0.003301704302430153, + 0.007489509414881468, + 0.012647606432437897, + 0.00538104772567749, + 0.0040818313136696815, + 0.002855398692190647, + 0.029717832803726196, + 0.01389158796519041, + 0.0027878540568053722, + 0.012003951705992222, + 0.009950892068445683, + 0.0067269145511090755, + 0.007673187181353569, + 0.009091923013329506, + 0.016257481649518013, + 0.020043283700942993, + 0.006604076363146305, + 0.011500205844640732, + 0.01953742839396, + 0.013524685986340046, + 0.0022317401599138975, + 0.0024059719871729612, + 0.004470429383218288, + 0.006370767951011658, + 0.001943552982993424, + 0.002540373941883445, + 0.0035627016332000494, + 0.0037631120067089796, + 0.0049720206297934055, + 0.003702771384268999, + 0.012653861194849014, + 0.008668579161167145, + 0.0058625005185604095, + 0.005473108030855656, + 0.013977698050439358, + 0.007496094796806574, + 0.005644171964377165, + 0.0043815248645842075, + 0.013227401301264763, + 0.00934139359742403, + 0.004928875248879194, + 0.008654986508190632, + 0.011618596501648426, + 0.010483426041901112, + 0.01292143203318119, + 0.010547108016908169, + 0.019821904599666595, + 0.010903975926339626, + 0.007734895683825016, + 0.008141526021063328, + 0.0113602876663208, + 0.01630559377372265, + 0.06601573526859283, + 0.06820911169052124, + 0.06172524392604828, + 0.05342874303460121, + 0.025402212515473366, + 0.02181331254541874, + 0.041391707956790924, + 0.028461409732699394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.004093126393854618 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file